Digital Twin Approaches for Gear NVH Optimization: A Literature Review of Modeling, Data Integration, and Validation Gaps

Abstract

Quiet drivetrains have become a central requirement in modern electric vehicles, where the absence of engine masking makes even subtle gear tones clearly audible. As a result, manufacturers are looking for more reliable ways to understand how design choices, manufacturing variability, and operating conditions shape gear noise and vibration. Digital Twin (DT) approaches—linking high-fidelity models with measured data throughout the product lifecycle—offer a potential route to achieve this, but their use in gear NVH is still emerging. This review examines recent work from the past decade on DT concepts applied to gears and drivetrain NVH, drawing together advances in simulation, metrology, sensing, and data exchange standards. The survey shows that several building blocks of an NVH-oriented twin already exist, yet they are rarely combined into an end-to-end workflow. Clear gaps remain. Current models still struggle with high-frequency behavior. Real-time operation is also limited. Manufacturing and test data are often disconnected from simulations. Validation practices lack consistent NVH metrics. Hybrid and surrogate modeling methods are used only to a limited extent. The sustainability benefits of reducing prototypes are rarely quantified. These gaps define the research directions needed to make DTs a practical tool for future gear NVH development. A research Gap Map is presented, categorizing these gaps and their impact. For each gap, we propose actionable future directions—from multiscale “hybrid twins” that merge test data with simulations, to benchmark datasets and standards for DT NVH validation. Closing these gaps will enable more reliable gear DTs that reduce development costs, improve acoustic quality, and support sustainable, data-driven NVH optimization.

1. Introduction

The electrification of automotive drivetrains has fundamentally altered the landscape of Noise, Vibration, and Harshness (NVH) engineering. In conventional internal combustion engine (ICE) vehicles, broadband engine noise often masks tonal transmission-related noise components. In contrast, electric vehicles (EVs) lack this masking effect, making previously negligible mechanical irregularities in gears audibly prominent [1]. As a result, tonal gear whine and vibration phenomena have become a major NVH concern in EV powertrain design [2].

Gear whine is mainly driven by Transmission Error (TE), the deviation between the driven gear’s actual angular position and its ideal kinematic position. TE originates from tooth deflection, surface waviness, and micro-geometry imperfections, which excite dynamic meshing forces at the tooth-passing frequency and its harmonics. These forces propagate through the shafts and housing, producing the characteristic tonal noise [3].

Even with today’s highly precise manufacturing, periodic deviations and elastic interactions can still generate “ghost orders”—subtle tonal components in the audible range that affect perceived sound quality and often escape traditional quality inspections [4].

While traditional TE mitigation strategies—such as macro-geometry optimization (e.g., helix and pressure angle tuning) and micro-geometry corrections—remain foundational, these deterministic methods are increasingly being enhanced with physics-based and data-driven approaches capable of modeling nonlinearities in time-varying mesh stiffness, damping, and lubrication conditions. Dynamic tribological and multibody simulation models demonstrate that even minor variations in lubricant stiffness or surface compliance can result in measurable changes in gear harmonic amplitudes [5].

The integration of optimization algorithms and simulation-based design—such as the use of genetic algorithms for gear modification—has demonstrated significant effectiveness in reducing TE fluctuations and enhancing tooth contact uniformity in electric vehicle (EV) transmissions. These approaches reflect a broader shift toward adaptive, predictive NVH modeling frameworks that account for real-world variations in load, manufacturing conditions, and operating environments [6].

Despite recent advances, most NVH simulation and optimization workflows still operate with fixed boundary conditions and static parameters, leaving them disconnected from real operational variability and unable to capture stochastic effects from manufacturing, lubrication, or loading.

Additionally, operational and end-of-line (EOL) testing data are rarely integrated back into design processes, maintaining a persistent disconnect between simulation and reality. Bridging this gap requires digital-twin-based frameworks that continuously synchronize virtual models with real-world behavior through real-time feedback, adaptive updates, and predictive control throughout the component lifecycle.

Digital Twin (DT) technology provides an integrated framework for addressing these challenges. A DT is commonly defined as a high-fidelity virtual representation of a physical asset that stays synchronized with its real counterpart through continuous life-cycle data exchange [7,8,9]. First introduced by Grieves in the early 2000s and later formalized by NASA in 2010 [9], the DT concept links real-time data, simulation models, and analytics to support prediction and system-level optimization [8]. In automotive applications, DTs connect multi-physics simulations with prototype or field measurements, creating a continuously updated model of the vehicle or subsystem. Siemens, for example, describes a vehicle-level NVH DT that “combines simulation and test data” throughout development. By fusing data from physical tests with CAE models, engineers can front-load NVH design and continuously update predictions as new data becomes available [10].

The goal of a DT is a closed-loop feedback system. Manufacturing variations and operating data update the simulation, while model predictions guide design changes or control actions [6,10].

Despite increasing attention, digital twin applications for gearbox and driveline NVH are still in an early stage. Most current implementations focus on predictive maintenance rather than direct acoustic or vibration optimization [11]. In gearbox design, advanced modeling tools such as Finite Element (FE), Multibody Dynamics (MBD), or boundary element (BEM) are still used mainly offline. These models rarely connect to live digital twin feedback loops, limiting real-time synchronization with the physical system. As a result, adaptive and predictive NVH control remains largely unrealized [12].

Developing a genuinely NVH-oriented Digital Twin involves several specific challenges. Such a system must capture high-frequency dynamics to reproduce tonal excitations, gear-mesh harmonics, and ghost orders, while also accommodating heterogeneous data sources ranging from end-of-line noise tests to vibration spectra and in situ acoustic measurements. These demands exceed the limits of traditional deterministic solvers, motivating hybrid solutions that blend physics-based and data-driven models [13]. Within this direction, multi-fidelity DT frameworks are gaining traction, enabling low- and high-fidelity models to operate jointly for real-time updating and uncertainty quantification [14].

However, a major barrier to DT adoption in NVH is the lack of standardized validation procedures. Current practice typically compares predicted and measured sound pressure levels or order spectra, yet no agreement exists on which metrics should define an acoustically reliable twin. Recent reviews of DT-based condition monitoring highlight that model fidelity, data integration, and uncertainty representation remain persistent bottlenecks [15].

Emerging research is now exploring real-time and lifecycle-aware model updating, where DTs are continuously adjusted to capture degradation and changing boundary conditions during operation. Machine-learning architectures such as Autoencoders and LSTM-based dynamic parameter tracking have been proposed to update digital models without requiring extensive offline experiments [16]. In parallel, acoustic applications of digital twinning are beginning to demonstrate closed-loop experimental optimization. For instance, Fushimi et al. [17] presented a digital twin approach for acoustic hologram tuning that integrates real-time microphone feedback and numerical differentiation, enabling rapid convergence of measured and simulated sound fields [17].

These developments suggest that bridging digital twin modeling, acoustic fidelity, and real-time data assimilation could open new possibilities for NVH applications. Yet, fully realizing an NVH-oriented DT will require harmonizing simulation accuracy, data fusion, and standardized validation—objectives that remain largely unresolved in the current literature.

This literature review aims to fill this knowledge gap by providing a comprehensive analysis of digital twin approaches for gear NVH optimization. The scope includes: (1) how gear dynamics and vibro-acoustic models have been integrated into DT frameworks; (2) how manufacturing data (geometry deviations, torque and vibration measurements, etc.) and IoT connectivity are used to update or inform these models; (3) the state of validation and benchmarking of NVH DTs; and (4) implications for sustainable engineering (e.g., reducing physical prototypes through virtual testing). Special attention is given to research gaps such as missing feedback loops between physical and virtual domains, underexplored use of machine learning alongside physics models, and the need for standardized metrics and frameworks [7].

Contribution and Novelty: To the authors’ knowledge, this is the first review explicitly centered on DT applications for gear and drivetrain NVH. Prior reviews of DTs in manufacturing or rotating machinery focus on reliability and maintenance; by contrast, we emphasize vibro-acoustic performance and design optimization. We not only synthesize the State-of-the-Art, but also cluster the open research gaps into a structured Gap Map (Modeling, Data Integration, Validation, Application/Sustainability), linking each gap to representative literature. We then propose concrete future research directions to address these gaps, serving as a roadmap for both researchers and industry practitioners [8].

2. Background and Definitions

2.1. Digital Twin Concepts and Standards in Manufacturing

The term DT refers to a continuously updated digital profile of a physical object, process, or system, maintained through data exchanged over its life cycle [18]. By dynamically reflecting the state of its physical counterpart, a DT enables real-time monitoring, simulation, and optimization within a virtual environment [19].

Grieves’ foundational framework distinguishes between a Digital Twin Prototype (DTP), used during design before any physical asset exists, and a Digital Twin Instance (DTI), a live virtual replica that updates throughout the asset’s operation. This distinction remains central and has since been expanded by subsequent work clarifying the roles of digital models, digital shadows, and digital twins within cyber-physical systems [20].

In manufacturing, DTs are commonly discussed alongside these related terms: a digital model is a static digital representation without real-time data linkage; a digital shadow allows one-way data flow from physical to digital; and a true DT implies bi-directional data exchange, where digital insights can trigger changes in the physical system [21].

Contemporary studies have emphasized the product lifecycle perspective of DTs, describing their evolution from design-oriented digital models to operational twins that support monitoring, predictive maintenance, and optimization across the full asset lifecycle [22]. Further domain-specific frameworks also propose clear taxonomies separating prototypes, shadows, and cyber-physical twins, aiming to standardize terminology across industries [23,24].

Standardization efforts for DTs are underway to ensure interoperability and common definitions. ISO 23247 (Automation systems and integration—DT framework for manufacturing) provides a high-level reference architecture for implementing DTs on factory floors [25]. It defines general principles for representing shop-floor entities as DTs and how they interact in an IIoT (Industrial Internet of Things) context. Another key enabler is the OPC Unified Architecture (OPC UA)—an open communication protocol widely adopted in Industry 4.0 to facilitate data exchange between heterogeneous devices and software. In the DT context, OPC UA enables real-time streaming of factory sensor data (vibrations, temperatures, etc.) into simulation models or databases [9]. For instance, an OPC UA server might aggregate a gearbox’s sensor readings, which the twin consumes to update its state. Another relevant standard is the Functional Mock-up Interface (FMI), an open standard for co-simulation and model exchange between tools. FMI allows different simulation components (e.g., a multibody drivetrain model and a control system model) to be linked as Functional Mock-up Units (FMUs) in a unified simulation [26,27]. This is crucial for building complex digital twins that span mechanical, electrical, and control domains of a vehicle. Although originally developed for system simulation, FMI 2.0 and the upcoming 3.0 facilitate integration of high-fidelity models into real-time capable simulations, which could help incorporate detailed NVH submodels into an overall powertrain DT.

It should be noted that as of 2025, there is still no NVH-specific DT standard or reference framework. The DT standards focus on general manufacturing and condition monitoring aspects, while NVH evaluation in industry follows separate standards (for example, ISO 1328 gear tolerances, or various ISO/SAE standards for vehicle noise testing). One gap identified is the lack of standardized data interfaces and metrics bridging these domains—e.g., a standard way to feed measured gear micro-geometry or noise levels into a digital simulation model. Recent industry initiatives have started addressing such integration: for example, the Industrial DT Association (IDTA) proposes an Asset Administration Shell concept for unifying asset data, and the DT Consortium’s liaisons aim to include domain-specific data conventions [28]. In the gear realm, some companies have developed proprietary data schemas for linking metrology data to simulation, but a common open approach remains elusive.

2.2. Gear NVH Fundamentals: Transmission Error and Tonal Orders

Understanding gear NVH requires a brief review of its primary excitation mechanisms and metrics. As introduced, TE is the main source of gear whine noise. TE is formally defined as the difference between the actual angular position of the driven gear and the position it would ideally be in if gear meshing were perfect, often expressed in arcseconds or micrometers. In essence, TE quantifies how much “slip” or deviation occurs in the gear pair’s motion—any nonzero TE means fluctuating mesh forces. Through Fourier analysis, TE can be decomposed into components at the gear mesh frequency (tooth passing frequency) and its harmonics, plus any sidebands [7]. For a spur or helical gear pair with N teeth running at rotation frequency f_rot, the mesh frequency is N × f_rot. TE oscillations at this frequency produce the fundamental gear mesh tone; higher harmonics (2×, 3× mesh frequency, etc.) produce superharmonic tones. These periodic forces are transmitted through shafts, bearings, and housing, causing structure-borne vibration and airborne noise.

In NVH terminology, ghost orders (or phantom tones) are another phenomenon of interest. These are tonal components at frequencies not integer-related to the mesh frequency (hence “non-synchronous” with the tooth mesh). Ghost orders often stem from periodic manufacturing deviations such as surface waviness or eccentricity in gears. For example, if every gear tooth has a slight sinusoidal waviness with a wavelength spanning a few teeth, it introduces a modulation in TE that appears as sideband tones (ghost frequencies) around the main mesh frequency. These can lead to annoying whining or humming noises, even if overall sound levels are low. Unlike mesh harmonics, which are an inherent result of tooth meshing stiffness variation, ghost orders are purely due to manufacturing or assembly errors. In practice, both mesh harmonics and ghost orders must be considered in a complete NVH analysis of a gearbox. Standard gear quality metrics historically focused on controlling transmission accuracy and runout, but did not explicitly quantify waviness or noise performance. This gap has motivated research into new measurement and specification approaches (for instance, defining a standardized “waviness spectrum” or metrics like TE RMS, peak-to-peak TE, etc., which correlate with noise [7]).

Finally, it is important to distinguish structure-borne vs. airborne NVH in gear systems. Structure-borne vibration refers to oscillations transmitted through the gearbox mounts, shafts, and vehicle frame—these may be felt or heard (after radiating via panels). Airborne noise is the acoustic pressure radiated by vibrating surfaces (gearbox casing, etc.) into the air. In many passenger vehicle cases, gear whine is first generated as structure-borne (the gear mesh excites the housing), which then radiates as airborne sound into the cabin. Simulating airborne noise requires coupling a structural dynamic model of the gearbox to an acoustic radiation model [10]. This can be computationally expensive. Often, NVH engineers use intermediate metrics as proxies, such as acceleration at certain points on the housing, or order tracks (filtering vibration or sound signals by order). For validation of DTs, common metrics include the amplitude of the gear mesh order (in dB) and its difference between simulation and test, overall SPL in a band, or even psychoacoustic metrics for tonal audibility. Section 5 will discuss how a lack of unified metrics hampers progress in NVH twin validation.

Although this review focuses primarily on gear and drivetrain excitations, vehicle-level NVH in electric vehicles is also strongly influenced by additional factors outside the gearbox. Suspension systems, subframe modes, and powertrain mounting dynamics can amplify or modulate structure-borne responses, particularly in the mid-frequency range. Aerodynamic sources become increasingly relevant at higher vehicle speeds and may mask, reinforce, or interact with tonal components originating from the drivetrain. These effects lie beyond the scope of the present section, but they form the broader NVH context in which drivetrain-focused digital twins must ultimately operate.

2.3. Multibody Dynamics, FE/BEM and Hybrid Modeling of Gear Trains

A variety of modeling approaches are used to simulate gear dynamics and acoustics. Multibody Dynamics (MBD) models represent the gearbox components (gears, shafts, bearings, housing) as masses and interconnected elements (springs, dampers, joints). Simple MBD models like the classical 6-DOF gear pair model capture torsional, translational, and rotational motions of two gears on shafts with a mesh stiffness element between them. This model is computationally lightweight and has been widely used to simulate gear vibration and TE under various conditions (it can predict resonances, basic sidebands, etc.). However, such simplicity sacrifices accuracy: it often assumes rigid shafts or bearings and cannot capture high-frequency modes or housing flexibility. More complex lumped-parameter models include additional coordinates for bearing deflections, gear rim bending, housing modes, etc., improving fidelity at the cost of complexity. MBD models are typically time-domain simulations solving nonlinear differential equations of motion (especially if including backlash or time-varying mesh stiffness). They are well-suited to study transient effects and nonlinear phenomena (e.g., tooth contact loss, rattle) [8].

In contrast, Finite Element Analysis (FEA) provides a spatially detailed model of components by discretizing them into elements. In NVH, FEA is often used to find the modal properties of a gear or housing (e.g., mode shapes and natural frequencies), which can then be coupled to an MBD model for a modal reduction approach [10]. Fully integrating FEA into a time-step simulation of gear contact is possible but computationally intensive; however, recent works have performed flexible multibody simulations with gear contact modeled via advanced contact mechanics [8]. For acoustics, the Boundary Element Method (BEM) or acoustic FEA is used to simulate how vibrating surfaces (like a gearbox casing) radiate sound. These frequency-domain methods can predict sound pressure levels at a microphone location given structural vibration data as input. Hybrid FE-BE approaches are standard for vehicle NVH (solve structure in FE, sound radiation in BE). In a DT context, including such high-fidelity models is challenging due to computational load—hence a gap exists in developing reduced-order models that retain acoustic-critical behavior but run faster.

Hybrid physics–data models are an emerging approach relevant to DTs. These include surrogate models and physics-informed ML (where training is guided by physics constraints). For gear NVH, examples include using machine learning to predict TE or noise based on manufacturing parameters. Horváth (2025) used Random Forest and neural networks to predict gear noise from features like surface waviness and profile errors [29], achieving good accuracy and thus acting as a fast surrogate for a detailed simulation. Another hybrid approach is model updating via measured data: a baseline physics model is corrected using test data to improve correlation. The concept of a hybrid physics-informed learner is illustrated in some frameworks, where high-resolution metrology and sensor data feed an ML model that is constrained or initialized by physics knowledge.

Overall, the modeling “toolbox” for gear NVH is rich, but each method has limitations in a DT context. High-fidelity FEA/BEM is accurate but slow; lumped models are fast but may miss important details. A key research thrust (discussed later) is how to achieve real-time capable yet accurate models—possibly by combining approaches (e.g., coarse MBD + ML corrector) or by model reduction techniques. Additionally, co-simulation standards like FMI allow linking these models, but ensuring stability and fidelity when coupling different domains (mechanical, acoustic, control) remains non-trivial.

It is also worth noting that meshless numerical methods have gained attention in recent years as alternatives to traditional FE/BEM discretization. Approaches such as Smoothed Particle Hydrodynamics, peridynamics, and other particle-based or mesh-free formulations can handle large deformations, crack initiation, and complex contact conditions without the burden of mesh generation. Although their application to gear NVH remains limited, these methods represent a promising direction for future high-fidelity digital twins, particularly where local stress–strain resolution and robust multi-physics coupling are required.

2.4. NVH Data Acquisition and IoT Integration

A DT for NVH relies not only on simulations but also on rich data from the physical world. In gear systems, relevant data sources span the product lifecycle:

• Manufacturing and Metrology Data: Modern gear production can provide detailed information on each gear’s geometry and quality. For example, 3D surface topography scans or single-flank roll-test results give profile error, lead error, pitch variations, and even waviness spectra for each gear [7]. These data could populate a “digital thread” from manufacturing to simulation [24]. Indeed, integrating measured micro-geometry into NVH simulation has been demonstrated: Romax Technology reported techniques to incorporate measured gear flank waviness into their NVH models [7]. Standards like ISO 1328 define [30] tolerance classes for gear deviations, but do not directly quantify expected noise. There is ongoing research into extending such standards or developing new ones to include acoustic performance indices. Additionally, process logs might indicate residual stresses or distortions that influence NVH; however, these are seldom used in current models.
• End-of-Line (EOL) Testing: Many transmission manufacturers perform EOL noise tests on gearboxes. This can involve running the gearbox on a test rig with microphones and accelerometers to detect anomalies (commonly known as noise grading or “gearbox whisper” tests) [7]. New EOL equipment integrates quality measurements: for instance, Gleason’s GRSL combines double-flank roll testing with laser scanning of the teeth, providing 100% inspection of TE and surface deviations in one machine. The data from such tests—essentially a short-run vibration spectrum and precise geometry—could be extremely valuable for a DT. If fed back into the design model, they allow last-minute model correction and also build a database for correlating manufacturing variables with NVH outcomes [29].
• Operational Sensors (IoT): In service, a running gearbox (especially in test fleets or prototype vehicles) may be instrumented with accelerometers, sound sensors, or torque/strain sensors. With the IIoT paradigm, these sensor streams can be sent via telemetry (e.g., telemetry units in test cars) or logged for later analysis. Cloud-based DT platforms (such as those by Siemens, GE, etc.) use protocols like OPC UA or MQTT to stream data securely to a cloud where the twin resides. For example, Hexagon’s cloud-based gearbox twin project uploads vehicle telematics from fleet vehicles to run simulations that predict fatigue life [24]; a similar approach could apply for NVH, uploading speed and load profiles and perhaps cabin noise measurements, to predict if a gearbox will develop a whine under certain driving patterns.
• NVH-Specific Measurements: Some advanced techniques measure NVH-related properties directly. Torsional vibration can be measured with optical encoders or laser torsion meters, yielding TE time histories in operation. Acoustic cameras or arrays can map noise sources on a running gearbox to identify dominant radiation spots. Order tracking with tachometer signals helps isolate gear orders from other noise sources in a vehicle test. These measured datasets are crucial for validating the DT’s predictions (e.g., comparing an order tracking plot from the twin vs. one from a physical test [7]. A gap, however, is that such rich data are not consistently stored or formatted to be reusable for modeling—often they are used ad hoc in troubleshooting.

The backbone enabling integration of all these data streams is the underlying connectivity and data infrastructure. OPC UA is widely recognized as a key interoperability standard in Industry 4.0, with many smart-factory implementations relying on OPC UA servers embedded in test rigs and machine tools, which the digital-twin platform can subscribe to [3]. Alternatively, for remote or distributed systems, lightweight IoT protocols like MQTT (Message Queuing Telemetry Transport) can push data to a cloud DT in real time. Indeed, Cho et al. [9] demonstrated a real-time synchronized DT factory using MQTT, noting that OPC UA ensured interoperability at the machine level while MQTT efficiently transmitted data to the central twin [9]. For NVH twins, one can imagine an MQTT topic streaming “gearbox1/encoder_TE” data that continuously updates the twin’s model of TE.

Efficient data management is crucial. High-frequency NVH measurements generate large datasets that must be stored and linked with model information, often using time-series databases such as InfluxDB [9]. Calibration details—such as test bench setups and environmental conditions—must also be tracked to ensure the digital twin remains consistent with real-world conditions.

Overall, while the sensors and IoT tools for gear NVH digital twins are already available, their integration remains inconsistent. Some companies are developing closed-loop NVH optimization workflows from design to operation, but standards are still lacking. Data exchange between design, manufacturing, testing, and field domains remains both a technical and organizational challenge. Later sections of this review highlight how weak data integration and missing feedback loops limit the effectiveness of NVH digital twins.

3. Methods of the Literature Review

This review follows a structured narrative approach based on established guidelines for technological reviews. It combines systematic database searches with expert insights to include both academic and industry perspectives. Major databases such as ScienceDirect, IEEE Xplore, Springer, MDPI, and Taylor & Francis were searched for publications from 2012 to 2025. Keywords were formed around DT and NVH in the context of gears and drivetrains, for example: “DT” AND (gearbox OR drivetrain OR transmission) AND (NVH OR noise OR vibration OR acoustic). Additional targeted queries included terms like “vibroacoustic simulation”, “transmission error”, “hybrid model”, “predictive maintenance”, combined with DT to ensure relevant works on condition monitoring were included if they had vibro-dynamics aspects. We also included specific searches for standards patents, white papers, and industrial reports (searching company websites and magazines such as Gear Technology, Gear Solutions, etc., for terms like DT noise gearbox).

Inclusion/Exclusion Criteria: From the search results, we screened titles and abstracts to include sources that dealt with vibro-acoustic modeling or analysis in gear systems or rotating machinery and referenced DT concepts (or closely related ideas like model updating, real-time simulation, IoT-based monitoring). We included peer-reviewed journal articles, conference papers, review papers, a selection of relevant patents, industry white papers, and a few standards/consortium reports. We gave priority to publications from 2016 onward to capture the recent DT wave, but also included a few seminal earlier works for background. We excluded papers that were purely about fault diagnosis or predictive maintenance unless they specifically incorporated a dynamic model or NVH aspect, as well as any studies with no connection to gear/drive systems. After initial filtering, ~120 sources were identified as potentially relevant.

Screening and Clustering: Each source was read in detail, and relevant content was extracted. Given the interdisciplinary nature (mechanical engineering, computer science, and standards), we organized the literature into thematic clusters: (1) Gear dynamics and NVH simulation methods; (2) DT frameworks and data integration techniques; (3) Applications of DTs in rotating machinery (including condition monitoring); (4) Validation, standards, and metrics. We also created a sub-cluster for sustainability and process efficiency impacts. While this is a narrative review (not a PRISMA systematic review of clinical trials), we followed PRISMA’s spirit for transparency; Figure 1 conceptually illustrates the flow, starting from ~300 initial search hits to 119 relevant sources used.

Quality and Bias: We critically evaluated each source’s credibility. Peer-reviewed sources from journals, including Mechanical Systems and Signal Processing, Journal of Sound and Vibration, IEEE Access, and Applied Acoustics, among others, were considered high-quality. Conference papers and theses were included if they filled a niche (some contain practical case data). Patents were not peer-reviewed; hence, we use them mainly to illustrate industry efforts or concepts rather than to rely on validated results. White papers and standards are cited to provide context. We attempted to avoid bias towards any particular software or company solution, focusing instead on the underlying techniques and challenges.

Through this method, we compiled 119 references that form the basis of the synthesis in Section 4, Section 5, Section 6 and Section 7. The information was synthesized qualitatively: we compared findings across studies to identify consensus, contradictions, and gaps. No new experimental data were produced, and any forward-looking analysis is based on patterns observed in the literature combined with our interpretation.

4. State-of-the-Art in Gear NVH Digital Twins

In this section, we review the current State-of-the-Art, organized into four aspects: (Section 4.1) Physical Measurements and Data Collection relevant to gear NVH twins, (Section 4.2) Simulation Modeling Techniques forming the backbone of the twin’s virtual model, (Section 4.3) Data Integration and Platforms enabling the twin (including IoT connectivity and standards), and (Section 4.4) Hybrid Approaches combining Physics and Machine Learning. For each, we highlight representative recent works and their capabilities, setting the stage for the gap analysis in Section 5.

4.1. NVH Measurements and Gear Data for Twins

The fidelity of a DT heavily depends on the quality and quantity of physical data available. In gear NVH, several measurement techniques generate data that can be leveraged:

• Transmission Error (TE) Testing: Single-flank testing is a standard method to measure TE under light load by running a gear against a master gear with encoders. Modern TE testers can output TE as a function of rotation angle, which can be converted to an order spectrum. Such spectra show the base mesh frequency and any sidebands (ghost orders). Some research test setups also measure TE under load to capture load-dependent mesh stiffness variation. TE data is invaluable for validating simulation models of the mesh; e.g., a study by He et al. (2014) combined analytical and FEA methods to compute TE and compared against measured TE, finding good correlation in lower orders but discrepancies at higher frequencies [31].
• Noise and Vibration Measurements: Gear whine is often evaluated by measuring sound pressure levels (SPL) near the gearbox or in the vehicle cabin. In laboratory setups, structure-borne noise can be assessed by accelerometers on the gearbox housing or nearby structures [7]. For example, a common end-of-line quality check is to have an accelerometer on a gearbox fixture while running gears at a specific RPM, then perform an FFT to identify tonal peaks [7]. A recent Gear Solutions article noted that increased use of accelerometers and even microphones in production testing is enabling the collection of baseline noise data for every gearbox [24]. Airborne noise testing may involve acoustic chambers or intensity probes; one challenge is that drivetrain noise in vehicles can be low in amplitude, requiring quiet environments or engine-off coastdown tests to measure.
• Optical Metrology and Waviness Detection: As introduced earlier, surface waviness on gear teeth is a subtle but important factor for NVH. Traditional gear inspection machines (CMMs, profile form testers) could not easily characterize waviness. New optical scanning methods (like laser deflectometry or 3D scanning) can capture the full surface topology of gear teeth. One 2023 patent by Klingelnberg (Gorgels and Finkeldey) describes an optical measuring system integrated with a rolling test to detect periodic surface deviations and correlate them with noise. This indicates an industry trend to directly link measured geometry to expected noise (creating what is essentially a DT element—the physical measurement feeding a noise prediction algorithm). Companies like Gleason have published white papers on laser-based waviness analysis for noise prediction, and standards bodies are discussing adding waviness parameters to gear standards. The availability of such high-resolution geometry data is a new opportunity—one study created a “digital twin of the gear” including its as-made micro-geometry and showed that incorporating actual measured topography into simulation changed the predicted vibration spectrum to better match tests [7].
• Torque and Load Sensors: To drive a digital twin (DT), accurate knowledge of the input excitation is vital. In test benches, torque transducers and rotational encoders record the driving torque and speed, which can then be used as boundary conditions for simulation. In vehicles, CAN-bus signals or inverter data often provide these inputs—especially in electric vehicles, where the motor controller can estimate instantaneous torque [32]. Advanced setups sometimes employ strain gauges or fiber-optic sensors mounted directly on shafts to capture dynamic torque fluctuations [33,34]. However, the complexity of signal transmission, sensor calibration, and thermal drift correction often limits their widespread use in production environments [35,36]. Nevertheless, incorporating measured torque time histories into NVH models greatly improves the accuracy of transient predictions, especially for phenomena such as gear rattle, load reversals, and shift clunk [37]. Future DT frameworks may further exploit telemetric torque data from in situ sensors to enable real-time correlation between measured loads and predicted acoustic responses [38].
• Operational Field Data: As DT concepts mature, operational field data is becoming increasingly central. In the wind energy sector, for example, DT systems rely on continuous vibration, temperature, and oil-quality monitoring from embedded sensors, with these data streams processed in cloud-based platforms to support gearbox condition monitoring and predictive maintenance [39,40]. In the automotive industry, continuous NVH monitoring in customer vehicles is not yet standard—mainly due to cost and the non-safety-critical nature of acoustic comfort—but fleet tests and durability prototypes are increasingly equipped with high-resolution NVH data acquisition systems [41]. Data collected during such campaigns is used both for correlation and model refinement of vehicle NVH digital twins. Recent work demonstrates that NVH data pipelines can be integrated into centralized cloud infrastructures, improving accessibility for design and validation teams. A few pioneering efforts have attempted to build full-vehicle NVH digital twins. For example, Tonelli et al. (2024) demonstrated a real-vehicle digital twin that continuously updates a multibody dynamics (MBD) model based on measured vibration data from an EV fleet [42]. Similarly, Prokop et al. [43] showed how experimentally measured gearbox vibration data could refine simplified NVH models for improved correlation [43].
• These examples illustrate that the identified gaps have quantifiable industrial impact, both in EV NVH development cycles and in high-volume gearbox production.

In summary, the measurement side is rich and becoming richer, but fragmentation is an issue: geometric data, vibration data, and sound data often live in separate silos. Bridging these in a cohesive way is part of the State-of-the-Art that only a few integrators have tackled. For instance, Romax/Hexagon’s work connects gear manufacturing data directly into fatigue life prediction in a digital twin; extending this to acoustic performance is a logical next step. The current State-of-the-Practice, however, largely treats NVH as a separate testing phase. Thus, one could say the data exists, but its use in DT frameworks is in early stages. The development of standardized “NVH databases” that link design, production, and test information for each gear unit could greatly accelerate DT accuracy. Efforts like the aforementioned VDI 2022 conference paper on digital thread for gear design-to-metrology highlight that the industry is aware of this need [24].

4.2. Simulation and Modeling Techniques for NVH Twins

At the core of any NVH DT is a simulation model capable of predicting the dynamic behavior and acoustic output of the gear system. We review the main modeling approaches currently used, noting how they contribute to a DT:

• Analytical and Lumped-Parameter Models: These include the classic torsional models and their extensions. The classical 6-DOF gear model is attractive for DT applications because it runs in real time, but—as Habbouche et al. [8] note—it cannot capture complex interactions or high-frequency effects [8]. More detailed models (e.g., 21- or 34-DOF) include bearing stiffness, shaft bending, housing flexibility, and mesh-stiffness variation from TE, offering much better prediction of gear-whine frequencies and amplitudes. Nevertheless, it uses linear assumptions (like constant damping) and neglects some nonlinear effects; therefore, its precision is moderate. The 34-DOF or higher models start approaching FE detail (including nonlinear bearings, localized fault simulation), making them suitable for detailed studies but heavy for routine twin operation. A DT might employ a hierarchical approach: use a simple model for continuous monitoring and a complex model offline for detailed analysis when needed. This concept of hierarchical twins (a fast twin and a slow twin) has been floated in the literature as a way to balance real-time needs with accuracy [8].
• Finite Element Models: Full 3D FE models of gearboxes capture modes of the housing, detailed tooth contact mechanics, etc. In NVH simulation practices, often a reduced FEM (modal model of housing) is coupled with a gear contact model. For example, one might derive a modal model of the gearbox casing (hundreds of modes up to, say, 5 kHz) and couple it to a lumped model of the gear pair via constraint modes at bearing interfaces [10]. This approach, a form of component mode synthesis, is implemented in tools like Romax NVH or Ansys. It yields a good prediction of resonance peaks and radiated noise when combined with acoustic BEM. Its drawback for DT: the models are large and usually require off-line precomputation; updating them in real time with new data is not straightforward. If a twin had to account for, e.g., a cracked rib in the housing discovered in inspection, updating an FE model is possible but not trivial. Reduced-order models (ROMs) such as response surface models or condensed matrices are one solution. Another approach is parametric FE models, where certain parameters (like material properties or boundary conditions) can be adjusted on the fly—some research exists on parametric modal analysis (for instance, to simulate how adding stiffness or mass changes frequencies, possibly relevant for updating a twin with retrofits).
• Multibody/Flexible-body Co-simulation: Modern MBD software (SIMPACK, MSC Adams, Simcenter 3D Motion, etc.) allows incorporating flexible bodies (from FE) and advanced gear contact models. The Simcenter 3D Motion “Transmission Builder”, for example, can include detailed gear mesh models with micro-geometry and calculate TE and forces, and these can be fed into acoustics solvers [10]. While such high-end simulations are mostly offline tools, Simcenter has started integrating them into a DT environment by coupling with test data in its Testlab software. Their Virtual Prototype Assembly (VPA) method can assemble measured FRFs of components (like a measured mounting stiffness) with simulated FRFs, essentially mixing test and simulation to create a hybrid model (which they refer to as the DT of the vehicle). This approach is State-of-the-Art in the sense of combining domains, but it is still largely a simulation/test correlation tool rather than a continuously updating twin [10].
• Surrogate and Data-Driven Models: On the emerging front, data-driven models for NVH are being explored. For example, machine-learning models have been trained to predict gear noise levels or spectra from design parameters. In one case, Sun et al. (2024) used support vector regression (SVR) to predict vehicle body sound insulation performance—not a gear case, but indicative of ML predicting acoustic metrics [44]. In gear context, Horváth [29] used Random Forest and XGBoost to predict which manufactured gears would be noisy, based on features like surface finish parameters [29]. These models achieved >90% classification accuracy in identifying noisy vs. quiet units in a production line, effectively serving as a “virtual noise sensor” in the production process. Such surrogates can be extremely fast (millisecond predictions) and thus suitable for real-time twin updates or inline quality control. However, their accuracy is bound by training data. If a twin encounters a scenario outside the ML model’s trained envelope (e.g., a new type of gear or an unusual wear pattern), predictions may be off. A hybrid approach keeps a physics model in the loop as a safeguard.
• Prognostics and Health Models: Another facet of simulation in DT is remaining useful life and fault modeling, often falling under predictive maintenance. While not NVH optimization per se, these models simulate fault progression (e.g., spalling, pitting on gear teeth) and their vibrational signatures. The wind turbine industry has many such models; e.g., Moghadam et al. (2021) developed a DT that uses a physics-based model of a floating wind turbine drivetrain and updates it with online condition monitoring data [11]. They could detect imbalance and misalignment by comparing the model-predicted vibration with the measured one. For automotive NVH twins, analogous health models could track if a gear’s noise is increasing over time, possibly indicating wear or lubrication issues. This crosses into maintenance, but in an EV where quiet operation is expected, detecting a rise in noise early is valuable.

In summary, the modeling State-of-the-Art is multi-faceted: high-fidelity models exist but struggle to be real-time; simplified models run fast but lack accuracy; and new ML surrogates are promising for speed but need physics awareness for reliability.

Table 1. Gear NVH Modeling Methods vs. Capabilities. (MBD = Multibody Dynamics; FE = Finite Element; BEM = Boundary Element; ROM = Reduced-Order/Surrogate Model).

Modeling Approach	Time/Frequency Domain and Scope	Real-Time Feasibility	Scalability (Model Size)	Validation Status (NVH)	References
Simple MBD (6-DOF, etc.)	Time-domain; gearpair torsion + 1D vib	Yes (ms time steps)	High (few DOF, low compute)	Qualitatively validated; misses HF details	[7,8]
Advanced MBD (≥21-DOF)	Time-domain; includes bearings, shafts	Borderline (depends on DOF)	Moderate (simplified flex)	Good for <2 kHz; some validation vs. tests	[7,29]
Flexible MBD (FE components)	Modal freq-domain + time integration	No (requires model reduction)	Low (hundreds of modes)	High fidelity; validated for modal behavior, not always for noise amplitude	[8,10]
FE Structural + BEM Acoustic	Frequency-domain; full 3D radiation	No (minutes/hours solve)	Low (thousands of nodes)	Accurate SPL prediction when inputs are correct; needs test correlation	[25,43,44]
Hybrid Physics-ML	Either domain; ML augmenting physics	Yes (ML is fast)	High (ML scales with data)	Early studies show good predictive accuracy, but limited scenarios have been validated	[11,12,13]
Pure ML Surrogates	Trained on past sims/tests (any domain)	Yes (very fast inference)	High (once trained)	Requires extensive validation; risk of extrapolation error	[31,32,33]
Prognostic/fault models	Time/freq; special cases (damage, etc.)	Varies (some real-time capable)	Moderate (per fault mode)	Validated for fault frequencies; such is less for noise loudness	[45,46]

summarizes the capabilities of key model types for gear NVH in a DT context. Notably, there is no one-size-fits-all—hence a trend toward hybrid modeling, using each model in contexts where it is the strongest. For example, an ML model might continuously estimate noise under varying conditions (fast), while a physics model runs in the background or periodically to ensure predictions remain grounded (accurate). The concept of DT fidelity levels is pertinent: one could maintain multiple twins at different fidelity and switch between them as needed.

Current practice in industry often uses a combination—e.g., a design stage uses FE-based modal analysis and acoustic prediction—while production might use simpler MBD or empirical models for pass/fail noise criteria. DTs aspire to unify these stages, but as seen, the computational and integration challenges are significant. The advances in computing (GPU acceleration, cloud parallelism) may help bring higher-fidelity models into near-real-time use in the coming years, which is a needed development for comprehensive NVH DTs.

Meshless and particle-based numerical methods are increasingly explored as alternatives to traditional FE/BEM solvers, especially for problems involving large deformations, contact nonlinearities, or crack initiation. Approaches such as SPH, peridynamics, and other mesh-free formulations offer improved numerical stability in scenarios where meshing is difficult or computationally expensive. While their use in gear NVH remains limited, these methods represent a promising direction for future high-fidelity digital twins that require robust multi-physics coupling and local stress–strain resolution.

Figure 2 summarizes this end-to-end architecture, showing how manufacturing metrology, TE measurement, simulation models, EOL NVH testing, and cloud-based data management form a continuous chain. Each stage enriches the next, ensuring that as-built deviations, system-level responses, and in-service feedback are captured in a single information flow.

4.3. Data Integration and Platforms

A DT is not just models and data in isolation—it requires an architecture to connect live data with simulations and to present outputs to users or other systems. Here we outline the platforms and integration methods currently seen:

• IIoT Platforms and Middleware: Many DT implementations leverage IoT platforms that handle data ingestion, processing, and user interface. Examples include Siemens MindSphere, GE Predix, PTC ThingWorx, etc. These platforms provide connectors to machines (via OPC UA, MQTT, etc.) and typically allow building analytics or simulation applications on top. In the NVH context, Siemens offers Simcenter Testlab integrated with MindSphere for remote monitoring of test data and feeding it to simulations [10]. Similarly, if one were to implement a gear NVH twin, an IoT hub could stream, say, gearbox vibration data from an EV fleet to a cloud, where a simulation (perhaps an FMU exported model) runs to predict noise or diagnose issues.
• Co-Simulation Frameworks: Functional Mock-up Interface (FMI) was mentioned; tools that support FMI can orchestrate co-simulations. For instance, a drivetrain model FMU could be fed with real motor torque data in a co-sim engine to simulate NVH in real time. Some research testbeds have used such setups to create “hardware-in-the-loop” simulations—essentially a form of DT where part of the system is real, and part is virtual. An example is a Ford project (2017) that co-simulated a multibody drivetrain model with a real engine on a dyno using FMI, to evaluate NVH and drivability concurrently [47]. The co-simulation ensured the engine and virtual transmission interacted realistically. This points towards future DT use in powertrain test cells, where part of the system could be replaced by its twin to expand testing capabilities.
• Data Models and Interoperability: A key challenge is ensuring interoperability between software and data formats. Some standards exist—ISO 10303 for gear geometry [48] and ISO 13336 for flank-form data [49]—but dynamic NVH data still relies on loosely defined formats such as ASAM ODS or CSV. As a result, most integration is custom: one OEM, for example, linked its gear metrology database to simulations via a Python API (Python v.3.9) that automatically inserts measured deviations into each DT instance. Reliable test–simulation merging also requires aligned reference frames and timestamps; consistent templates, coordinate systems, and naming conventions—as emphasized in the Simcenter workflow (Siemens Digital Industries Software, Simcenter Testlab v.2023.1)—are essential for robust DT data integration [10].
• Feedback Loops and Control: Beyond just one-way monitoring, advanced DTs implement control feedback. For NVH, an example might be an active control system using the twin’s prediction. Consider an active noise cancellation (ANC) wherein the twin predicts that a gear whine tone will exceed a threshold in the next second based on current operating conditions; hence, it triggers a countermeasure (maybe adjusting motor torque slightly to shift the frequency or activating a noise canceling signal through speakers). While such an application is quite advanced and not reported in the literature yet, the building blocks exist: fast predictions (via surrogate model) and integration with control systems (via vehicle CAN or similar). In industrial machines, there are instances of DTs used for dynamic tolerance adjustment—Rajkumar et al. (2025) describe an AI-DT architecture where measured deviations trigger adjustments in the machining process in real time [50]. Translating that to NVH, one could imagine in manufacturing, if a gear’s measured waviness predicts noise issues, the twin could recommend re-grinding or pairing that gear with a better mate in assembly.
• User Interface and Visualization: NVH engineers typically use waterfall plots, Campbell diagrams, and color maps to visualize noise and vibration. DT dashboards must convey such information intuitively. Some tools provide “auralization”—synthesizing the predicted sound, such that engineers can literally listen to the twin. This is already performed with NVH simulators, where you can listen to how a car would sound with different components. Integrating that into a DT means if the twin’s parameters update (say a gear wear increases whine by 5 dB), a product engineer could immediately hear the effect. This can significantly aid decision-making (sound quality is hard to judge by numbers alone). HEAD Acoustics, in a white paper, described an “NVH Simulator: A path to the DT”, highlighting how virtual acoustic prototypes can be combined with physical measurements to allow subjective evaluation early. This shows State-of-the-Art tools moving toward immersive DT experiences [51].

The conceptual data flow from the physical system to the predictive Digital Twin core is summarized in Figure 3, illustrating the typical NVH-oriented integration pipeline.

In summary, the State-of-the-Art in integration is characterized by silo-breaking attempts. Testing, CAE, and operations are coming together, but most efforts are bespoke. A few standards (OPC UA, FMI) provide a backbone, but on top of that, each implementation might choose a different data architecture (central cloud vs. edge computing, etc.). For gear NVH specifically, there is not yet a dedicated commercial “NVH twin” platform. Instead, companies piece together existing tools—e.g., using Simcenter or LMS Test.Lab for data acquisition and analysis, Romax or Ansys for simulation, and an IIoT platform for connectivity. The vision is that eventually a unified platform could handle all of it, but interoperability remains a stepping stone being actively worked on [9].

Figure 4 provides a concise overview of how NVH-relevant information moves through the digital thread—from metrology and simulation inputs to end-of-line results and in-service feedback. By linking these previously isolated data sources through a common AAS-based architecture, the workflow enables consistent model validation and faster root-cause identification. The result is a closed, continuously improving NVH development loop across the drivetrain’s entire life cycle.

4.4. Hybrid Physics–ML and Emerging Approaches

One of the most exciting trends in recent years, as identified in multiple sources, is the convergence of traditional physics-based modeling with machine learning in the context of DTs. We have touched on this throughout, but here we compile the State-of-the-Art approaches and examples:

• Physics-Informed Machine Learning (PIML): In PIML, the ML model is trained or constrained using physical laws. An example relevant to gear NVH is using simulated data (from many runs of a physics model with varied parameters) to train an ML model, but also enforcing known physics in the model structure. Liu et al. [52] conducted a systematic review of DT components and highlighted PIML as a key to maintaining twin fidelity [52]. In gear applications, one could use a limited set of expensive FEA acoustic simulations to train a neural network that predicts noise spectra from input parameters (module, teeth number, misalignments, etc.), ensuring the network respects symmetry and scaling laws known from physics. There is also research on neural nets that take vibration signal inputs and output diagnostic classification, with a physical model in parallel that generates synthetic signals for training.
• Reduced-Order Models and Surrogates: Surrogate modeling is a well-trod approach in NVH for optimization. Techniques like response surface methodology or polynomial chaos expansions have been used to create algebraic models of NVH metrics as functions of design variables. The new twist is using them in an updating loop. For instance, if a surrogate is embedded in a twin to predict TE given current temperature and load, it can be continuously corrected by bias factors derived from any measured TE. An applied example: Kobayashi and Alam (2024) created an explainable AI model for predicting remaining useful life in a DT, focusing on making it trustworthy for engineers [53]. This reflects the need in industry to have ML models whose outputs can be explained in physical terms (e.g., “Noise increased because surface roughness exceeded X, as per the model”).
• Federated and Transfer Learning: These approaches are emerging as promising approaches for intelligent diagnostics in scenarios where data is scarce or distributed across multiple sites. In gear NVH analysis, for example, imagine multiple factories each manufacturing similar gears: a federated learning (FL) scheme could train a global NVH prediction model without requiring any factory to share its raw data, thereby preserving privacy and proprietary information [45]. Combining FL with transfer learning (TL) enables the adaptation of global models trained on one machine, component, or product to new gearboxes with limited local data [46]. These approaches have already proven effective in cross-device fault diagnosis, enabling collaboration between separate facilities and even different equipment types without compromising data confidentiality [54]. In NVH and acoustic applications, transfer learning allows deep models trained on one signal type—such as vibration—to be adapted for another, like airborne acoustic data, improving the model’s versatility across sensor modalities. Although not yet widespread in gear NVH, these methods are expected to become increasingly relevant as digital twins and data-driven gear diagnostics expand across distributed manufacturing ecosystems [55].
• Edge Computing for Twins: A challenge for real-time twins is latency in sending data to the cloud and back. Some propose performing more at the “edge” (on a local device or test rig). For NVH, an edge device could run a simplified model or inference of an ML model on an FPGA or local PC, such that immediate feedback is possible. Habbouche et al. (2025) discuss deploying “compressed DT models near the monitoring stations” to improve real-time detection on wind turbines [8]. In automotive testing, one could similarly run a twin on a high-performance laptop connected to the test vehicle, rather than relying on cloud processing.
• Integration of Diverse Data (Multimodal): NVH does not exist in a vacuum; operating conditions, environmental data (temperature affects noise via lubrication viscosity, etc.), and even subjective feedback could be incorporated. Some cutting-edge approaches attempt to fuse different data types. For instance, one might combine vibration signals with high-speed camera footage of the gear mesh (it is possible with stroboscopic methods) to directly correlate tooth deflection patterns with noise. While no production system does this yet, research like You et al. (2024) used an intelligent fusion of sound and vibration signals via an AI model for bearing fault diagnosis [56]. Extending to gear NVH, combining airborne and structure-borne sensors through AI could yield more robust detection of emerging noise issues than either alone.
• Standards and Frameworks with AI: On the standards front, the ISO 23247 [57] series formally defines a reference architecture for digital twins in manufacturing and explicitly accommodates AI-enabled analytics and interoperability [58]. The upcoming ISO 23247-4 component is expected to further specify guidelines for analytics within twin systems. Recent reviews have emphasized how AI-driven predictive maintenance frameworks rely on digital twins for continuous learning and fault prediction, yet still face barriers such as poor data quality, sparse fault labels, and the need for lifecycle model retraining [59]. Together, these developments indicate a growing convergence between AI standards and digital twin frameworks, highlighting the need for standardized architectures that support adaptive and intelligent twin evolution.

In summary, the State-of-the-Art is that hybrid DTs for gear NVH are on the horizon, with initial examples appearing in the literature. By 2024–2025, the field would have clearly shifted from pure physics models towards combined approaches.

The reviewed studies reveal converging trends in how NVH-oriented digital twins are structured. Most architectures integrate three essential layers: the physical system with embedded sensing, the digital representation based on numerical and analytical models, and the control/optimization loop. These relationships are synthesized in Figure 5, which depicts the common flow of information and feedback found across multiple studies.

The framework highlights the bidirectional data exchange between the Physical System and the DT, establishing a continuous synchronization loop. Manufacturing and metrology data define the baseline model, while sensor feedback and simulation updates refine it dynamically. The control and optimization layer ensures the system adapts to operational changes, achieving TE reduction and NVH improvement.

The concrete benefits of these hybrid approaches in NVH are improved predictive accuracy (especially for phenomena hard to model, such as ghost orders or friction noise) and adaptive learning (the twin can “become smarter” as more data comes in). However, they also introduce new challenges: trust and interpretability of ML components, data governance, and maintaining physical consistency. These will appear again in the gap analysis (Section 5), as the community has started identifying open issues around hybrid twins.

4.5. The Role of the Digital Thread in Gear Twin Architectures

While the concept of the DT focuses on creating a high-fidelity virtual representation of the physical system, the digital thread serves as the underlying data backbone that ensures continuity and traceability across the entire product lifecycle.

For NVH-oriented gear systems, the digital thread connects design models, manufacturing data, and operational measurements into a consistent information flow. This continuity allows engineers to trace noise and vibration behavior back to design parameters or process variations, thus supporting data-driven decision-making and continuous improvement.

As illustrated in Figure 6, the digital thread spans all major lifecycle stages—from design and manufacturing to operation, maintenance, and end-of-life. Each stage contributes unique datasets, all of which feed back into the design phase.

The figure illustrates the continuous flow of information connecting the design, manufacturing, operation, maintenance, and end-of-life phases, enabling traceability and feedback throughout the product lifecycle. This closed-loop flow forms the foundation for an integrated DT ecosystem, enabling not only predictive NVH analysis but also sustainability gains through recycling and design optimization based on in-field feedback.

4.6. Bearing Dynamics and Their Influence on Gear NVH

Bearings play a key role in the vibration and noise behavior NVH of gear systems because they directly affect gear-mesh dynamics. The stiffness of rolling-element bearings is nonlinear—it depends on the current load, deflection, and local contact conditions. As torque passes through the gears, the number of rolling elements carrying load changes, which causes time-varying bearing stiffness and sometimes sudden stiffness jumps when new elements enter or leave the load zone [60]. These fluctuations influence the dynamic response of the gear pair because each bearing revolution includes several mesh cycles that generate sidebands and tonal components in the vibration and noise spectra [61]. According to Hertzian contact theory, stiffness increases with load, while under light loads, partial contact or internal clearance effects may occur. If bearing clearance becomes too large, rolling elements can accelerate and decelerate through the load zone, sometimes impacting the cage and producing vibrations at the fundamental train frequency [62]. Conversely, applying a preload removes clearance and increases stiffness, improving system stability but shifting resonance frequencies higher [63]. Designers must balance these effects to avoid “looseness” vibrations from excessive clearance or overly stiff behavior from high preload. Thermal effects also change bearing stiffness: at high temperatures, radial stiffness decreases while axial stiffness increases, leading to nonlinear dynamic behavior such as periodic or chaotic motion in gear systems [64].

Bearing deflections under load (and due to thermal effects) alter gear alignment and mesh. At high speeds or temperatures, a bearing may deform: for instance, thermal expansion can change radial and axial stiffness. One study on an EV helical gear drive found that increasing speed and temperature raised bearing axial stiffness but reduced radial stiffness, leading to nonlinear gear dynamics. Including the effect of thermal deformation in the bearing model caused the gear system to exhibit rich nonlinear behavior (periodic and even chaotic responses) as speed changed. The authors conclude that bearing stiffness changes must be considered to accurately predict high-speed gear NVH. In general, bearing load–deflection curves are nonlinear; hence, using a constant stiffness can mispredict gear vibration amplitudes. A recent 2025 study introduced a displacement-dependent bearing stiffness model in a 6-DOF gear dynamics simulation, showing that incorporating time-varying bearing stiffness significantly alters dynamic results, increasing vibration amplitude compared to a constant-stiffness assumption [64].

Another key factor influencing gear NVH behavior is bearing support flexibility—the compliance of shafts and the gearbox housing. When shafts or housings deform under load, the bearing mounts can move dynamically, worsening TE fluctuations. Studies have shown that housing elasticity amplifies vibration levels. For instance, coupling housing deformation into a gear–bearing model increased vibration amplitude by about 60% and nearly doubled bearing reaction forces [65]. This occurs because housing deflection effectively reduces bearing support stiffness, allowing more gear misalignment and variation in mesh forces. Similarly, flexible shafts introduce additional bending vibrations, shifting natural frequencies, and modifying resonance modes of the gearbox [66]. Dynamic models show that when housing or shaft stiffness decreases, the system’s natural frequencies (especially bending and torsional modes) move closer to the operating range, amplifying gear noise and vibration [67]. In extreme cases, resonance can occur within the normal speed range, causing loud tonal components. Bearing resonance phenomena such as cage whirl or rolling-element passage frequencies may also couple with housing flexibility, further modulating vibration transmission through the system [68]. Therefore, both shaft and housing stiffness are critical design parameters: rigid supports help isolate gears from structural vibrations, while compliant supports can amplify them if not properly managed.

If these coincide or modulate the gear mesh frequency, distinct sidebands appear in noise spectra. For example, a slight oscillation of the shaft in the bearing can modulate gear meshing forces and produce sideband tones known as “ghost frequencies.” Bearing fundamental train frequency can modulate gear noise if clearance allows micro-oscillation; excessive clearance was shown to cause large cage-pocket impact forces at this frequency [69]. These modulations are perceived as annoying tonal beats or roughness in the sound.

Bearing surface imperfections also contribute. Waviness on bearing races—a form of manufacturing error—creates periodic variations in contact force. This can impose an additional excitation on the gear shaft. A recent study confirmed that bearing waviness greatly influences gear vibrations: higher waviness order or amplitude intensified overall vibration levels, and an uneven waviness distribution led to even larger vibrations than perfectly sinusoidal waviness [70]. In essence, bearing waviness can modulate gear mesh vibrations, producing sideband frequencies around the gear mesh order. These are similar to the effects of gear tooth waviness, but originating in the supports. Such waviness-induced modulation is known to contribute to tonal “ghost noise” in transmissions [70]. Another subtle effect is micro-slip in bearings—minute slips or creeps between rolling elements and raceways under oscillatory motion. Micro-slip can introduce damping in some regimes, but if combined with frictional stick-slip, it might excite resonances or cause fretting damage (leading to noise). For instance, vibration experts note that micro-slip under small oscillations can cause fretting wear and noise in stored bearings. In operation, micro-slip at the rolling contacts could generate high-frequency bursts of vibration due to momentary slippage. Although less documented in the literature than stiffness and waviness effects, micro-slip is a potential NVH issue, especially in scenarios with oscillatory loads or during start-stop cycles [71].

To predict and mitigate these influences, modern NVH simulations include detailed bearing models. In multibody dynamics (MBD) models of gear systems, bearings are often represented by nonlinear spring-damper elements. These are computed using Hertzian contact theory to capture load-dependent stiffness. For example, an engineering case study reports using Hertzian contact formulas to calculate bearing stiffness, accounting for localized deformation at each rolling element contact [72]. The model can update stiffness at each time step based on the instantaneous load direction and magnitude. Damping in the bearings (from lubricant and material damping) is also included to represent energy dissipation. Some advanced MBD tools allow detailed rolling element representation or incorporate the bearing supplier’s data for stiffness vs. load. Clearance and preload are modeled by allowing a small free range before stiffness engages, or by initial compression of the spring. This enables simulation of impacts if a loose bearing’s clearance is suddenly taken up, or the stiffening effect of preloads on vibratory response.

In finite element (FE) or flexible body simulations, bearings can be included via condensed stiffness matrices or contact elements. A common approach for housing or shaft analysis is to include the bearing as a set of spring elements with nonlinear properties derived from curves. Some researchers integrate bearing contact analysis directly into FE models—e.g., modeling the outer ring flexibility and housing bore with a contact interface to simulate how housing deformation changes bearing stiffness [65]. This coupled approach reveals interactions between gear vibrations and bearing support deformation. Additionally, to capture bearing eigenmodes, detailed FE models of the bearing components can be coupled to the system. However, this is computationally expensive and often simplified in system-level NVH models.

In summary, gear NVH cannot be analyzed by considering the gears alone—the bearing support dynamics are fundamental in shaping the overall noise behavior. Modern NVH studies treat the gearbox as an integrated gear–bearing–housing system, where the nonlinear stiffness, damping, and clearance of bearings play a critical role in modulating gear vibration responses. Bearings introduce load-dependent stiffness variations and can display nonlinear and time-varying behavior due to Hertzian contact effects, thermal deformation, or preload changes. These effects generate sidebands and complex spectral patterns, which can either amplify or suppress vibration amplitudes depending on operating conditions. Research between 2020 and 2025 consistently shows that ignoring bearing nonlinearities or simplistically modeling stiffness as constant can significantly underestimate vibration levels or miss key resonant peaks. Advanced models that include variable bearing stiffness and damping have demonstrated more accurate correlation with experimental vibration spectra, capturing phenomena such as chaotic motion, sideband modulation, and “ghost tones” under high-speed conditions. Incorporating these effects into both MBD and FE models—through Hertzian contact springs, parametric stiffness formulations, and high-fidelity FE modeling of bearings—leads to far more realistic NVH predictions. Such methods also reveal how preload, housing flexibility, and bearing surface errors affect vibration patterns and TE modulation. This integrated understanding enables practical NVH mitigation—for example, defining optimal preload to prevent clearance-induced impacts, tightening bearing surface tolerances to suppress waviness excitation, and stiffening housings to shift structural resonances beyond operating frequencies. Consequently, bearing–gear interaction stands out as a key determinant of gearbox quietness and operational smoothness.

4.7. Industry 5.0 Digital Twin Extensions

The transition from Industry 4.0 to Industry 5.0 represents a fundamental shift from a technology-driven model focused on automation and efficiency toward a more human-centric, sustainable, and resilient paradigm. While Industry 4.0 prioritized cyber-physical systems, IoT connectivity, and real-time optimization, Industry 5.0 expands this vision by placing humans at the center of industrial ecosystems and ensuring that technology serves people rather than the reverse [73]. In this new context, digital twin (DT) frameworks have evolved from mirroring machine operations to modeling human–machine interaction, trustworthiness, and ethical transparency. Research highlights that Industry 5.0 digital twins aim to incorporate human digital twins (HDTs)—digital replicas of workers that capture physiological, cognitive, and behavioral data to improve well-being, safety, and collaboration [74]. New frameworks also embed ethical principles such as fairness, privacy, and transparency into system design (“trust by design”), ensuring that AI and DT-based decision-making remains accountable and explainable [75]. This ensures workers remain “in the loop” and that technology supports human creativity and decision-making rather than replacing them [76]. Ultimately, Industry 5.0 envisions technology working with and for people—merging the digital intelligence of machines with human insight and adaptability. Digital twins are thus being reimagined as socio-technical ecosystems that balance efficiency, ethics, and empathy, aligning industrial progress with societal and environmental well-being [77].

One of the most notable developments within Industry 5.0 is the emergence of Human-Centric Digital Twins (HCDTs). While earlier digital twins primarily mirrored machines or production processes, HCDTs explicitly model human operators, their interactions, and even their physical and cognitive states within the system [74]. A Human Digital Twin (HDT) extends conventional digital human models by integrating real-time biosensor and motion data—capturing indicators such as ergonomics, fatigue, or stress—and combining these with AI-driven analytics. This enables real-time monitoring of both machine performance and worker well-being, supporting proactive interventions (e.g., adjusting workstation layout or robot behavior to reduce strain) [78,79]. Recent studies highlight that human-centric digital twins use technologies such as motion capture, biosensors, AR/VR, and cognitive modeling to enable safer, more intuitive, and more adaptive human–machine collaboration [80,81]. For example, AR-integrated digital twins allow operators to visualize machine states and receive adaptive guidance, while VR-based HCDTs can simulate ergonomic stress or task performance in virtual environments before real-world implementation. In practice, an assembly line digital twin might adjust robotic assistance levels based on a worker’s real-time physiological feedback, reducing fatigue and improving safety. Similarly, HDTs in manufacturing and construction settings are being used to monitor emotional and mental states through multimodal data—such as speech, facial expression, or heart rate—to enhance safety and reduce human error [82]. New frameworks, such as the “Perspectives–Observer–Transparency” paradigm, explicitly emphasize modeling the human in a transparent, interpretable way that respects privacy and cognitive limits [83].

Modern frameworks incorporate security, privacy, and explainability as core requirements. For instance, the Digital Twin Consortium’s Security and Trustworthiness working group notes that a twin operating on faulty or manipulated data can erode trust and cause bad decisions. They advocate for shared responsibility security models, given that a DT is a system-of-systems. Concretely, trust-based architecture might use blockchain or secure data provenance to ensure the data feeding the twin has not been tampered with and implement user access controls. Transparency entails logging the twins’ decisions and making their logic interpretable to stakeholders. A recent framework (“Digital Twin V”) was introduced to guide building trusted twins, stressing verification of models and clear documentation of assumptions [84]. In Industry 5.0, a twin is not just a technical tool but part of a socio-technical system where users must have confidence in it. As such, frameworks now include layers for ethics and governance—for example, a manufacturing DT might include modules to enforce data anonymization for worker data, or to explain why a predictive maintenance twin is recommending a shutdown (so that engineers can validate the reasoning).

A cornerstone of future-proof DT frameworks is the Asset Administration Shell (AAS). Originating in Industrie 4.0, the AAS is essentially a standardized digital wrapper for any asset, containing all its relevant data in a structured form. The AAS concept has gained traction moving into Industry 5.0 as a way to ensure interoperability and transparency. The AAS provides a common language for assets—whether a machine, a sensor, or even a human—to ensure that different systems and organizations can share digital twin data consistently. Importantly, AAS has been extended to incorporate human elements in Industry 5.0. In late 2023, researchers demonstrated a Human-AAS (H-AAS) that augments the standard asset shell with submodels describing human operator characteristics. This includes classes for worker attributes, emotional or medical state, and more. By conducting such, the human becomes an integrated part of the digital twin framework (not an external factor). The role of AAS in Industry 5.0 is to facilitate semantic interoperability—it ensures that when data is exchanged, the context and meaning are preserved. The AAS also acts as a container for trust and transparency; as it is standardized, any stakeholder can read an asset’s shell to understand its capabilities, data sources, and even how its data was computed. For instance, a twin’s AAS might include a submodel detailing the AI algorithms used and their training data—providing transparency. The use of AAS also supports vendor-neutral ecosystems since it is an open standard, multiple companies’ devices and software can interoperate, which is key for collaborative Industry 5.0 environments [85].

Another major development in digital twin (DT) technology is the integration of cyber-physical systems (CPS) with cloud-edge computing. In Industry 4.0, most digital twins operate either on local servers or entirely in the cloud to monitor and optimize machines. However, Industry 5.0 introduces a hybrid cloud–edge approach, where computing tasks are distributed between local and cloud environments. In this model, critical real-time functions and privacy-sensitive data stay on the edge—close to the production floor—while large-scale analytics and long-term optimization are handled in the cloud [86]. This setup ensures that local systems remain operational even when the cloud connection is lost, improving resilience and reliability. It also increases trust, since sensitive operational data can stay within the local network or be sent to the cloud only after aggregation or anonymization [87]. For example, in an automotive NVH testing scenario, an edge device could analyze real-time vibration signals to detect anomalies and stop defective units instantly, while summary data is sent to the cloud for trend analysis across multiple production sites. This collaborative edge–cloud model allows computations to occur where they are most efficient, reducing latency and optimizing energy use [88]. By 2025, such architectures are expected to include local “edge twins” for every critical asset—connected directly to sensors for fast updates—and a cloud twin that integrates data from all assets for factory-level analytics [89]. This approach also supports human-centric goals: edge devices can host AR/VR interfaces with low latency for on-site workers, while cloud systems allow remote experts to visualize, interact with, and guide operations in real time [90].

A central goal of Industry 5.0 DT frameworks is to ensure that advanced technology supports human and societal well-being. Alongside human-centricity, sustainability has become a core principle. Modern DT architectures are therefore designed to reduce energy consumption, optimize resource usage, and provide transparent sustainability metrics, such as real-time carbon footprint tracking [91]. The combination of the Asset Administration Shell (AAS) and OPC Unified Architecture (OPC UA) exemplifies this sustainable, interoperable approach. OPC UA provides real-time, secure communication between machines, while the AAS gives that data a semantic context, describing what it means and how it should be interpreted [92]. Together, they enable both speed and meaning in data exchange—an essential capability for sustainability analytics and transparent operations. For instance, an AAS submodel can store sustainability data, such as a product’s carbon footprint, which is computed using live energy data streamed through OPC UA connections. A 2023 implementation demonstrated how such data can be shared securely across digital twin data spaces, allowing partners to trace and verify carbon information in the manufacturing chain [93]. This ensures that sustainability metrics are not just numbers but traceable, reusable data entities within the digital twin ecosystem. In addition, the AAS–OPC UA integration enhances interoperability and security, ensuring trust and data integrity across systems. Recent work also shows how the AAS can manage secure provisioning of OPC UA applications, making industrial data sharing both automated and auditable [94]. Beyond interoperability, co-simulation standards such as the Functional Mockup Interface (FMI) are gaining traction in Industry 5.0. These enable different simulation tools to exchange models and run joint analyses in real time. For example, FMI-based setups have been used to combine vehicle drivetrain, control algorithms, and NVH simulations into unified studies—allowing engineers to evaluate how software changes affect mechanical vibration and noise. Open frameworks integrating AAS, OPC UA, and FMI are now emerging as reference architectures for sustainable manufacturing, supporting both model transparency and vendor-neutral collaboration [95]. Together, these technologies embody the Industry 5.0 vision: open, explainable, and sustainable digital ecosystems that link machine performance, human insight, and environmental responsibility.

Overall, Industry 5.0 pushes digital twins to be more holistic—merging technical and human systems—and governed by values of sustainability, usability, and trust. The latest publications (2023–2025) showcase frameworks where digital twins are not isolated “virtual machines” but parts of a collaborative, human-aware architecture for the factory and product lifecycle. These principles are directly feeding into next-generation automotive NVH digital twins, ensuring that as we digitally replicate and predict drive-train behavior, we perform it in a way that aligns with human goals (comfort, safety), uses data responsibly, and fosters cross-domain innovation.

4.8. AI-Enhanced Digital Twins and the Evolution Toward Industry 5.0

Recent research emphasizes that Artificial Intelligence (AI) plays a central role in advancing Digital Twin (DT) technology within Industry 5.0, transforming it into a more human-centric, sustainable, and intelligent ecosystem. Zhang et al. (2025) propose an AI-enhanced DT systems engineering framework that integrates the industrial metaverse with human–machine collaboration, enabling adaptive decision-making and ethical data governance in cyber-physical systems [25]. Similarly, Massaro [96] provides a systematic review of AI-driven DTs, highlighting their capacity to achieve trustworthy automation, cognitive manufacturing, and human-in-the-loop interaction through embedded intelligence. In parallel, Fedak et al. (2023) demonstrate the use of DTs for smart manufacturing optimization, showing that real-time DT integration can minimize energy consumption, reduce production defects, and improve equipment resilience—foundational elements for sustainable Industry 5.0 facilities [97]. Barata and Kayser (2024) extend this view, arguing that digital twins will be the “nervous system” of Industry 5.0, orchestrating collaboration between humans, machines, and digital ecosystems through interoperability and semantic standardization [73].

Finally, Isaza Domínguez (2024), in a systematic literature review, synthesizes Industry 5.0 DT research trends and concludes that the next generation of digital twins must embed ethics, transparency, and social sustainability at the core of their design [98]. Together, these studies converge on a shared vision: the AI-empowered digital twin as the enabler of adaptive, ethical, and sustainable Industry 5.0 ecosystems, where technological intelligence augments—not replaces—human expertise.

5. Identified Research Gaps

Despite the progress outlined in Section 4, significant gaps and challenges remain before DTs can be fully realized for gear NVH optimization. Through our literature analysis, we have clustered the research gaps into four broad categories: (Section 5.1) Modeling Gaps, (Section 5.2) Data Integration Gaps, (Section 5.3) Validation and Metrics Gaps, and (Section 5.4) Application and Sustainability Gaps. Each category is discussed below, with specific gaps, their implications, and supporting references.

5.1. Modeling Gaps in NVH Digital Twins

Gap M1: Limited fidelity of real-time capable models. One of the foremost gaps is achieving high modeling fidelity while retaining real-time or near-real-time performance. Current high-fidelity gear NVH models (FE-based or detailed flexible multibody) are too slow for an interactive twin, whereas real-time capable models (6-DOF, analytical) miss important dynamics [8]. This gap is echoed by multiple sources. Horvath’s review on gear surface waviness notes that while DT approaches have been proposed, their validation is “limited” and presumably one reason is that models in those DTs are oversimplified [7]. Similarly, Habbouche et al. (2025) stress the need for model reduction: “a key challenge is balancing model fidelity and computational load in DTs”, especially for capturing broad frequency ranges in wind turbine gearboxes [8]. Another aspect is multi-scale modeling—gear noise involves tooth-level phenomena (micron-scale errors) and system-level effects (meter-scale housing modes). No existing twin seamlessly spans those scales; most ignore the micro-scale or treat its effects empirically.

Gap M2: Insufficient modeling of manufacturing variations and their NVH effects. Traditional NVH models assume nominal geometry, perhaps with some “fudge factors” for damping or misalignments. However, as discussed, manufacturing variations like surface waviness, heat treatment distortion, and component tolerances can significantly affect NVH. There is a gap in integrating as-manufactured data into models. While some works have shown this concept (e.g., integrating measured waviness into simulation), they are not common. ISO gear tolerances do not directly map to noise; a DT needs a way to take measurement data (profile error charts, lead variation, runout values) and modify the model accordingly. The gap includes the lack of widely accepted model updating techniques for these variations. One patent by Elder [99] proposed adding a tailored compliance to gear teeth (via micro-geometry) to reduce TE and noise—essentially a design method to mitigate manufacturing issues. Such a patent suggests that the industry felt a need to solve noise by design tweaks rather than modeling, because modeling the random deviations each gear has was difficult. As a result, many gear NVH simulations still assume “perfect” gears and are then surprised by noise in production [29]. The impact is that without this, a twin might confidently predict low noise while the real gearbox (with minor tooth waviness) screams. Future research must focus on geometry-informed NVH modeling: how to efficiently morph a gear mesh model to include measured deviations, how to include stochastic variation in a fleet of twins. Some initial research uses statistical energy analysis or Monte Carlo on gear misalignments to predict variance in noise, but more is needed to directly tie measurement to NVH outcome in the twin [7].

Gap M3: Lack of multi-physics coupling (thermal, lubrication, etc.). Gear noise is influenced by factors like lubrication (which affects damping and contact stiffness) and temperature (which changes clearances). Most NVH models in the literature consider structural dynamics only, with maybe a simple representation of damping. A true twin should reflect all relevant physics. For instance, if gear oil temperature rises, viscosity drops, damping might reduce, and whine might worsen. Or consider EVs: motor electromagnetic noise can couple with gear noise. Currently, such multi-physics coupling in a twin is rare. One multi-physics mention is in an IEEE Access 2023 study that developed a hybrid model-based twin for drivetrain life prediction, including lubrication conditions, Ref. [8], but not specifically NVH. Another is an older Peking Univ. study [100] which attempted a co-simulation of gear dynamics with a thermal network to predict dynamic TE, including thermal expansion—these kinds of integrated models are not mainstream. The gap is partly due to complexity and partly due to siloed expertise. The result is that a DT might predict noise correctly at 20 °C, but if an end-user drives the car hard and the gearbox heats to 80 °C, the twin might no longer be accurate. Similarly, electric motor control strategies (like PWM frequencies) can modulate the noise—a twin that does not include the motor’s torsional excitation might misattribute noise sources. Bridging this gap might leverage co-simulation frameworks and modular twins—for example, linking a thermo-elastic model of the gearbox (to achieve deformations and mesh backlash changes) with the NVH model. This area is still open for significant work.

Gap M4: Representing complex systems and interactions (e.g., full vehicle). Most current research on gear twins focuses on a single gearbox or component in isolation. However, NVH issues often arise from interactions: gear resonance exciting a chassis boom, or a motor control causing gear noise modulation. A system-level NVH twin is largely absent. EU’s ECO-Drive project (2020–24) explicitly targeted “system-level NVH optimization for electric drivetrains” [101], implying it is still a work in progress. Creating a twin of an entire vehicle’s NVH (including gears, motor, road noise, etc.) is extremely complex, but could yield huge benefits in early design. The Simcenter approach of Virtual Prototype Assembly is a step in this direction, but that is still an engineering tool rather than an automated twin [102]. The gap is integrating subsystems—often they are modeled by different teams using different tools.

Industrial Example—EV gearbox TE-model mismatch: In a 2024 EV reducer development program (Tier-1 supplier, anonymized), TE prediction errors of 20–30% at the first mesh harmonic resulted in three additional design iterations. Each iteration required new prototype gears, adding approximately two weeks of lead time and an estimated 12–18 k€ per loop.

Root-cause analysis later confirmed that the mismatch originated from ignoring load-dependent bearing stiffness in the baseline MBD model.

In summary, Modeling Gaps center on fidelity vs. speed trade-offs, handling real-world variability, incorporating multi-physics, and scaling from component to system. Without addressing these, DTs for NVH will remain either too simplistic to add value or too slow to be practical.

5.2. Data Integration Gaps

Gap D1: Incomplete feedback loop between manufacturing, testing, and simulation. Perhaps the most frequently mentioned gap is the lack of closed-loop data flow in current practice. Many DT visions talk about a feedback loop: data from manufacturing and operation feedback to update the model, which then informs design adjustments or maintenance. In reality, these loops are only partially implemented. For instance, a gear manufacturer might measure all gears, but that data is rarely used to update the NVH predictions of the gearbox—it is used for QA and then mostly archived. Horváth and Feszty’s review explicitly identified “missing feedback loops among manufacturing data, measurements, and simulation updates” as a research gap [7]. Likewise, in condition monitoring literature, Zhang et al. (2023) note that while DTs are conceptually closed-loop, “most applications are one-directional (monitoring) without active feedback” [33]. The impact is suboptimal models and missed opportunities—e.g., a simulation might assume a nominal bearing stiffness—but if EOL testing showed that a particular unit has a slightly loose bearing preload, the twin never knows; thus, its predictions are off. Additionally, field data might show a trend, but without feeding that back, the design cannot improve. Closing this loop is non-trivial—it requires not just technical connectivity but trust in the data [8]. Some preliminary work exists, like an update method for multi-dimensional DT models by Zhang et al. (2023) [33].

Gap D2: Lack of standardized data formats and interoperability for NVH data. OPC UA provides a pipe but not a data schema for NVH. There is no widely adopted “gear DT data model” or ontology that defines, say, what a “transmission error signal” or a “gear acoustic order” is in a machine-readable way. This lack of standardization makes DT integration costly, as each new sensor or simulation tool often requires a custom adapter. NIST highlights that OPC UA and STEP can help, but both need extensions for emerging DT use cases. A clear example is that gear inspection machines still export proprietary formats, requiring conversion scripts before CAE use. Standards such as QIF or STEP AP242 extensions could address this, but no consensus exists. Similarly, test data stored in ASAM ODS or MDF formats cannot be directly imported into most simulation tools without manual preprocessing. The impact of no standards is slower adoption and error-prone integration. One company’s twin solution might not port to another because the data assumptions differ. Addressing this likely involves collaborative efforts to define a data exchange format for NVH-related DTs. This could tie in with existing standards: for example, extending FMI to support not just co-simulation but also streaming of test data, or using ASAM’s OpenX (which covers driving scenarios) to also include acoustic data [103,104].

Gap D3: Limited use of real operating data in design-phase simulation (and vice versa). This is related to D1 but from a life-cycle perspective: The design phase uses simulated data to predict NVH and might build prototypes for testing; the operational phase (vehicles in the field) gathers data via telematics or service, but these two phases are seldom connected in the NVH context. In maintenance-oriented twins (like wind turbine gearboxes), field data is used to plan repairs, but seldom loops back to redesign. For automotive NVH, rarely would a manufacturer feed back field NVH data (like customer complaints or recorded sounds) to update the design simulations of the next model, at least not systematically. This is a gap because one promise of DTs is continuous improvement: the twin accumulates knowledge over a fleet and informs next-gen designs. A 2021 review by Kooning et al. noted the potential of DTs in wind energy, but found that “closing the loop to design is still largely aspirational” [105]. In automotive, warranty databases contain NVH issues, but linking that back to design parameters requires manual forensic work. A twin that remains linked after sale could, in theory, correlate usage patterns to noise degradation and suggest design tweaks. The gap is both cultural and technical. The impact is that designs might repeat mistakes or be over-conservative because that feedback is not formalized. Bridging this gap may involve employing AI on field data to detect patterns and feeding those into design simulations. For example, if many cars show a certain bearing wear causing noise at 100k miles, the twin could simulate that wear and validate if a design change (like a harder bearing) would avoid it. Hardly any published work directly demonstrates this closed-loop for NVH; conducting such would be novel.

Gap D4: Cybersecurity and data ownership concerns are hindering integration. An often overlooked aspect: connecting physical assets to digital systems introduces security and IP concerns. In a factory, sharing gear measurement data to a cloud twin might be resisted if the data is sensitive. Or in an automotive context, streaming high-rate NVH data from vehicles raises bandwidth and privacy questions. Technically, solutions like data encryption, on-premises processing, or synthetic data generation can mitigate this, but research is needed on how to make sure a DT platform is secure and that companies feel safe using it for critical NVH decisions. While not highlighted heavily in the literature, some consortium reports note “data governance and security” as a barrier to full DT adoption [8].

Industrial Example—Missing manufacturing–simulation feedback loop: A 2023 production issue at a European gearbox plant (OEM anonymized) showed that unmodeled tooth waviness in a single batch increased EOL tonal failures by ~4%.

The lack of automated data transfer from CMM to the NVH simulation workflow delayed root-cause identification by almost two weeks, causing rework and scrap costs exceeding 20 k€. A digital-thread-based metrology → simulation link would have prevented the delay.

In summary, Data Integration Gaps are as much about process and standards as technology: the need to seamlessly connect all life stages of the gear (design, production, test, operation) in data terms. Until this is solved, DTs will operate on partial information, undermining their value. This category of gaps is somewhat easier to address with collective action (standard bodies, agreements); however, this requires impetus, which might come if a few success stories demonstrate clear ROI for closing the loop.

5.3. Validation and Metrics Gaps

Gap V1: Lack of standardized validation metrics for NVH digital twins. A consistent theme is that there is no consensus on how to measure the accuracy or effectiveness of a DT in the NVH domain [7]. For physical testing, we have standardized metrics, but how do you validate a twin? Is it by comparing simulation vs. test SPL at certain points? Or comparing order traces? The Horváth waviness review explicitly identifies “lack of standardized … metrics for consistent comparison across studies” as a gap. Without common metrics, each researcher or company might claim their twin is “accurate” based on different criteria, making it hard to benchmark progress. Potential metrics could include: TE RMS error (between predicted and measured), Order amplitude difference in dB for key orders, Correlation coefficient between predicted and measured spectra, coherence between model and test responses, etc. Another aspect is computational performance metrics—e.g., update latency. None of these is standardized. The impact is that the development of NVH DTs is somewhat ad hoc; a paper might show a correlation plot for one case and call it a day, while another might use a different approach. To build confidence, especially in industry, one needs clear acceptance criteria. Establishing such metrics might borrow from existing NVH and CAE correlation standards—e.g., there are guidelines for model correlation in vibration, or perhaps from ASME standards on verification and validation of simulation models. This gap suggests future work to propose an NVH-DT validation protocol, perhaps via a group like SAE or ISO.

Gap V2: Limited validation against experimental data in the literature. Many papers propose frameworks or models, but have minimal experimental validation. For example, in DT papers, one often sees a case study on a simplified lab setup or sometimes a purely simulated case (digital twin tested on “virtual data”). For NVH, real-world validation is crucial due to the many uncertainties. A 2025 review on gear waviness noted “limited validation of DT approaches against measured data” [7]. Habbouche’s wind turbine DT review similarly found very few instances of implemented twins validated with field measurements of gearboxes [8]. One reason is the practical difficulty—setting up a twin and running parallel physical tests is costly and data-heavy. However, without it, we do not know if these methods actually work in the field or if they just fit the few cases they were tuned for. The gap is in the knowledge of robustness and accuracy in diverse conditions. For example, a twin might work for one gearbox design tested in a lab, but does it work for a different design or in a vehicle? The gap is partly due to a lack of common test platforms. One idea is to establish a benchmark dataset or testbed for gear NVH DT research—e.g., an open dataset of a gear transmission with known geometry, measured TE and noise across loads, etc.—such that researchers can test their twin approaches comparably. Some initial efforts in sharing have occurred in general prognostics, but not for NVH specifically. Impact: Without broad validation, the industry will be hesitant to trust DTs for NVH-critical decisions. Therefore, filling this gap is essential for adoption.

Gap V3: Difficulty in validating subjective sound quality aspects. NVH is not just about raw dB or vibration; how humans perceive the noise (tonality, annoyance) is key. A DT might predict a certain sound spectrum, but how to validate that its predictions align with subjective impressions? There is a gap in linking twin outputs to sound quality metrics and validating those. Some work has been performed on objective metrics for gear whine annoyance. However, a twin would need to incorporate those and be validated against jury evaluations—i.e., if the twin says gear whine tonality index is X, do people indeed rate that noise as annoying? This is seldom addressed. The HEAD Acoustics white paper suggests using NVH simulators for auralization in a twin [51], but that is more of a design tool. If a manufacturer aims to use a twin to ensure sound quality, they must validate not only physics but perception—a high bar. Right now, this is a gap.

Gap V4: Unclear validation of hybrid models (physics + ML). When a DT uses machine learning, validating it becomes tricky because ML can sometimes give the right answers for the wrong reasons. Traditional validation is necessary but perhaps not sufficient; one would also want to validate that the model is physically plausible and will hold in new scenarios. Some authors like Kobayashi (2024) emphasize “explainable and trustworthy AI” in the twin [53]. This is a gap in methodology: how to validate that a twin’s ML component is not overfitting or will remain valid under slight extrapolations. Potential ways are cross-validation with different operating conditions, or ensuring ML features have physical meaning. But a formal framework is lacking.

Gap V5: Validation in operational environment vs. lab. Often, models are validated in controlled lab conditions. However, in operation, boundary conditions differ (temperature, mounting stiffness in cars, etc.). A twin might match the lab test but fail in the car. Bridging that gap could involve testing in vehicles. Not much literature exists for the full vehicle NVH twin validation; this is more of a future aspiration.

The consequence of these validation gaps is a credibility gap. Without proven validation, DTs for NVH remain academic exercises or pilot projects. Industry adoption will require demonstrating that a twin can reliably replace or augment physical tests. This ties into Gap 5.4 as well—proving the benefits.

A direct example: Suppose an automaker wants to reduce prototypes by using a DT. To justify removing a physical test, they would need evidence that the twin predictions fall within acceptable error bounds. Currently, there is no standard “twin certification.” Perhaps in the future, standards or regulators might allow simulation evidence if validated to X%.

5.4. Application and Sustainability Gaps

Gap A1: Unquantified impact of DT on the development process and sustainability (saving cost, time, and material). One of the touted benefits of DTs is the reduction in the need for prototypes and physical testing. However, as of 2025, there are few concrete case studies quantifying this for NVH. How many prototypes can be cut? What is the CO₂ or cost saving? This gap is partly because full adoption has not yet occurred to measure, and partly because we lack methodologies to credibly estimate these benefits in advance. The literature on sustainability is often qualitative. For example, Horváth’s predictive modeling work mentions “supports sustainable manufacturing by minimizing prototype production and resource consumption”, but no numbers are given [29]. The European Commission has funded projects (such as ECO-Drive) in part to achieve greater “virtual prototyping”, but the results are not yet public. Without hard data, management may not invest in NVH twins. Therefore, a gap is developing in a framework to evaluate the life-cycle benefits—e.g., a model that, if you implement a twin at such-and-such fidelity, you can skip a certain noise test or reach market Y months faster. The impact of closing this gap would be an easier justification for twin projects—aligning with corporate sustainability goals (prototypes and physical tests consume lots of energy and materials, as does iteration, aftermarket launch if NVH problems are found late). Perhaps the gap could be addressed through pilot programs that document metrics or through academic case comparisons.

Gap A2: Underdeveloped use-cases beyond monitoring (e.g., active noise control, adaptive systems). Most current DT deployments revolve around condition monitoring or predictive maintenance—essentially passive observation and prediction. In NVH, DTs could enable active use cases—e.g., an adaptive control system that adjusts a gearbox actuator or an active mount in real time based on twin predictions—or enable “what-if” scenarios to be run quickly during a test. These advanced applications are seldom demonstrated. They require not only the twin model but also integration with control actuators or design-optimization loops. The gap is that the potential of DT is not fully exploited—we mostly see monitoring, but not much on optimizing NVH in a closed-loop way. One reason is that adjusting NVH often requires design changes that cannot be performed on the fly. However, in EVs, some aspects, such as motor torque profiles or active acoustic noise cancellation, could be adjusted. A concept might be a twin that predicts a tonal noise in a scenario and pre-emptively adjusts the control to avoid it—an example of predictive control using the twin. We did not find explicit papers on that. It might be under exploration in OEM R&D but not published. The gap is a bit futuristic but relevant as systems become drive-by-wire and such.

Gap A3: Human factors and trust in using DTs for NVH decisions. Engineers are used to physical testing for NVH sign-off. Bringing them (and their managers) to trust a DT’s output for a critical decision is a gap that is more psychological/organizational. It will require extensive validation (Gap 5.3), but also possibly new ways to present results (like showing confidence intervals, explaining predictions).

Gap A4: Economic and business-case analysis for NVH twins. Companies often ask about the ROI of developing and maintaining an NVH-focused DT, yet such analyses are rarely reported. Estimating ROI requires balancing development, data infrastructure, and lifecycle maintenance costs against savings from reduced testing, fewer NVH-related warranty issues, and shorter time to market. Because many DT projects remain in pilot phases, comprehensive business-case studies are largely missing. Without clear economic evidence, large-scale industrial adoption remains slow, linking this gap both to sustainability (Gap A1) and to financial feasibility. Some white papers (like by Hexagon/Romax [106]) tout how a twin can bring insight, but they do not provide numbers like “we saved X dollars by avoiding Y prototypes”.

Gap A5: Regulatory acceptance and standard procedures. If a twin is to replace some testing, for things like vehicle homologation, regulators would have to accept simulation. This is beyond research, but it is part of the application gap. Thus far, pass-by noise still requires physical tests, although companies heavily simulate it beforehand. Perhaps in the future, a certified DT approach could streamline compliance testing, but currently, there is no pathway for that.

Gap A6: Training and skills gap. Implementing DTs requires skills in simulation, data science, and IoT—skillsets that traditionally reside in different roles. Many NVH engineers are experts in testing and CAE, but not as much in IT/IoT or ML.

Section 6 collates the main gaps (M1–M4, D1–D4, V1–V5, A1–A6) and indicates where the community should focus research efforts to unlock the full potential of DT technology for gear NVH.

6. Research Gap Map

To provide a consolidated view, we present a Research Gap Map summarizing the key gaps identified in Section 5.

Table 2. Research Gap Map—Categories, Specific Gaps, Impacts, and References.

Category	Gap Description	Impact/Why It Matters	Representative References
Modeling (M)	M1. Real-time vs. Fidelity Trade-off: High-fidelity NVH models are too slow; fast models lack accuracy.	Limits the twin’s ability to predict NVH accurately in real time, reducing trust in twin outputs.	[7,8]
	M2. Manufacturing Variability in Models: Twin models rarely incorporate actual gear deviations, surface finish, etc.	Twin may mis-predict noise if as-built differences (waviness, misalignments) significantly affect NVH (common in practice).	[7]
	M3. Multi-physics Coupling Lacking: Thermal, lubrication, and motor dynamics are not integrated into NVH models.	Missed phenomena (e.g., noise changes with temperature, or EM torque ripple) could lead to wrong predictions under certain conditions.	No explicit ref (inferred from multiple sources, e.g., [29] on factors beyond geometry)
	M4. System-Level Modeling: NVH twins usually focus on one gearbox, not full vehicle or multi-component interactions.	Unable to capture NVH issues arising from system interactions (e.g., gear noise amplified by vehicle body modes). Suboptimal holistic optimization.	[7,10] (implying need for system-level approach)
Data Integration (D)	D1. Broken Feedback Loops: Manufacturing and test data not fed back to update models (one-way flow only).	Twin does not evolve or correct itself with new data—leads to divergence from reality over time; missed continuous improvement.	[7,25]
	D2. Lack of Standard Data Formats: Different stages (design, CAE, test, IoT) use incompatible data schemas; no standard NVH twin data model.	High integration effort, risk of errors; hampers collaboration and tool interoperability—each twin implementation is bespoke.	[9] (OPC UA widely adopted, but there is a need for a schema)
	D3. Limited Use of Field Data in Design: Operational NVH data (from fleet/field) seldom loops to influence design or calibration.	Loss of valuable insights—designs might not address certain in-service NVH issues discovered later, or be overly conservative.	[8] (DTs not yet used for design feedback)
	D4. Data Security and Ownership Issues: Concerns over sharing sensitive manufacturing/test data on digital platforms.	Companies hesitant to fully integrate data, leading to partial twins; potential data silos persist due to IP/security risks.	Mentioned in the context of industrial DT challenges (e.g., DTC reports)
Validation and Metrics (V)	V1. No Standard Twin Validation Metrics: No agreed criteria (error bounds, correlation indices, etc.) to validate NVH twin accuracy.	Difficult to compare methods or certify a twin for use; reduced confidence and no clear targets for improvement.	[7]
	V2. Sparse Experimental Validation: Many proposed twins or models lack thorough validation against physical test data, especially under varied conditions.	Unproven reliability—risk that twins work only for demo cases. Industry adoption slowed by lack of evidence on generality/accuracy.	[7] (notes limited validation in the literature); [8] (calls for more experimental studies)
	V3. Ignoring Sound Quality Metrics: Twins validated on engineering metrics (dB, TE) but not on human perceptual metrics (tonality, annoyance).	Twin might meet numeric targets, but still result in customer-noticeable noise issues (tonal whine). Could misprioritize fixes.	Implied in [8] (importance of linking to comfort), HEAD acoustics whitepaper
	V4. Validation of Hybrid Models: Hard to validate ML-driven parts of twin beyond the dataset; lack methods to ensure ML does not unpredictably fail.	Limits trust in AI-enhanced twins. Without explainability and robust validation, engineers may not rely on twin recommendations.	[29] (need for interpretable, trustworthy AI in DT)
	V5. Operational Validation: Twin not validated in a real operational environment (e.g., in-vehicle), only in the lab.	Twin might not account for factors present in the field (mounting flexibility, background noise). Performance in real use remains unknown.	No specific ref; general industry practice gap
Application and Sustainability (A)	A1. Unquantified Benefits: Lack of data on time/cost/prototype reduction achieved via NVH twins.	Management unsure of ROI—hesitant to invest. Sustainability claims remain qualitative, slowing corporate buy-in for twin initiatives.	[29] (claims benefits, but no quantification)
	A2. Underutilization in Active Control/Optimization: DTs mostly used for monitoring/analysis, not yet for real-time control or automated design optimization loops in NVH.	Missing opportunity for DTs to directly reduce NVH (e.g., adaptive noise control, on-the-fly tuning) or to speed up design via automated twin-driven optimization.	Emerging concept, e.g., predictive maintenance vs. predictive control gap
	A3. Human Trust and Adoption: Engineers and decision-makers may be skeptical of twin outputs, especially if counterintuitive or from “black-box” models.	Risk of DT not actually being used, or always overridden by physical tests “just in case,” negating benefits. A cultural/educational gap could stall implementation.	[8] (DT as Industry 5.0 enabler requires trust/human integration)
	A4. Skills and Workflow Integration: Implementing NVH DT needs cross-domain skills (IT, data science, CAE); current workflows are siloed.	Slower development and deployment of DTs; potential errors when domain knowledge is not integrated (e.g., data scientists vs. NVH experts). Need training and possibly new roles/tools.	Discussed in Industry 4.0 workforce context, e.g., Jerman et al. 2020 on competencies [107]
	A5. Standards/Regulations Acceptance: No official guidelines to use DT results for certification (noise regulations, etc.).	Even if the twin is accurate, companies cannot replace certain physical tests due to regulatory mandates. Slows full utilization of DT.	Implied by the existence of strict test standards—no alternative methods defined yet

categorizes the gaps into Modeling, Data Integration, Validation, and Application/Sustainability, and highlights the impact of each gap along with representative references from the literature. This serves as a quick reference “map” for researchers and practitioners to identify where efforts are needed.

Looking at this map, one can see that the challenges are both technical (modeling, data) and organizational (validation practices, human factors). The next section (Future Directions) will address each category of gaps with potential research and development directions that can close these gaps.

7. Future Directions and Recommendations

Addressing the above research gaps will require concerted efforts from academia, industry, and standards organizations. In this section, we outline future research directions and actionable recommendations for each gap category (Modeling, Data, Validation, Application). These suggestions are drawn from trends in the literature and our analysis of what is needed to advance the State-of-the-Art.

7.1. Advancing Modeling Fidelity and Speed

Future Directions—Modeling 1 (FD–M1): Develop multi-fidelity and reduced-order modeling strategies. To tackle the real-time vs. fidelity trade-off (Gap M1), one promising approach is a multi-fidelity twin—a concept where the DT can operate in different modes or levels of detail depending on the context. For example, a coarse model could run continuously for fast updates, while a fine model runs in the background or on demand. Research should explore algorithms for seamlessly switching between model fidelities or blending their results. Reduced-order models (ROMs) derived from FEA could capture most of the NVH behavior with far fewer degrees of freedom [8]. Recent work in other domains (aerospace) shows success using machine learning to create ROM surrogates of FE simulations; applying this to gear whine (training an ML model on many FEA runs to predict noise) is a concrete future direction. Hints of this appear in Rajkumar et al. (2025), where an AI-DT architecture is used for predictive maintenance—similar could be performed for NVH: an AI that quickly predicts TE or SPL, constrained by physics-based model outputs [50]. The goal should be to achieve at least an order of magnitude speed-up in simulation without more than, say, 10% loss in accuracy for key metrics. Success here would directly enable real-time NVH twin usage [50].

Future Directions—Modeling 2 (FD–M2): Integrate manufacturing data into simulation parameterization. For Gap M2, we recommend creating methodologies to input measured gear data (micro-geometry, tolerance deviations) into simulation models automatically. This could involve extending gear contact models to accept tooth profile maps as input—some software already allows a sort of profile error map import, but it is not common. Another approach is statistical: develop models that predict how certain measured deviations (like a 5 µm waviness of order 10) will alter TE and noise [7], and update simulation parameters accordingly. The twin could have a library of “influencers”—e.g., profile error -> how much it increases first-order TE, etc., based on prior simulation or test results. A very concrete step would be to use 3D surface scan data of a gear to directly compute a custom excitation signal for an MBD model. Research along these lines is already in a Romax technical note, which tried incorporating measured waviness [7], but that knowledge should be generalized and published. Ultimately, digital thread integration should be demonstrated: a gear is made, measured, its DT model is updated with those measurements, and the predicted NVH is compared to test—closing the loop for one piece. Performing this for many samples would prove viability. If successful, future gear design could include a step: feed prototype manufacturing data into the twin to predict if it will meet NVH targets before physical testing is performed, enabling earlier issue detection.

Future Directions—Modeling 3 (FD–M3): Multi-physics coupling research and standardized interfaces. To fill Gap M3, research should extend gear DT models to include thermal and lubrication aspects. This could mean co-simulating a thermal network or CFD-based lubricant model with the structural model. A near-term way is using existing tools like coupling Romax with GT-Suite or Amesim (for thermal/lubrication) via co-simulation. But more academically, one could derive simplified equations linking temperature to mesh stiffness and damping and include those in the twin. Machine learning might also help—e.g., a model can be trained to predict NVH changes due to temperature or oil properties. Another key is including the electric motor dynamics for EV drivetrain NVH—for instance, incorporate motor torque ripple spectra as an input to the gear dynamic model. The functional mock-up interface (FMI) standard could be leveraged to couple different domain simulators; future standards might need to include semantics for linking these (addressing part of Gap D2 as well). A practical recommendation is to demonstrate a case study of a thermo-mechanical-acoustic DT: maybe an experimental rig where gear temperature is varied and the twin is able to predict the shift in whine frequency or amplitude due to thermal expansion or oil thinning. Proving that multi-physics integration yields better accuracy would justify the added complexity. In the long run, such integration would allow twins to predict issues like “after 20 min of driving, when the oil is hot, gear whine increases by 3 dB,” which currently purely mechanical models cannot do.

Future Directions—Modeling 4 (FD–M4): Hierarchical and modular modeling frameworks. For Gap M4 (system-level modeling), a future direction is creating hierarchical twins that can zoom in/out between subsystem and system. For example, a vehicle NVH twin might have modules for the gear set, the motor, the driveline, and the vehicle body. Using component-based models (as in Siemens’ VPA approach) and standardized interfaces like FMUs for each can allow assembly of a full system model. Research should address how to maintain accuracy when connecting these modules—e.g., how to ensure the coupling between gearbox vibrations and vehicle acoustics is properly represented. One idea: use transfer path analysis (TPA) results to create coupling elements between subsystems in the twin. Markus Brandstetter’s Simcenter blog suggests using Frequency-Based Substructuring to assemble a full vehicle NVH model; academics could further develop the theory of performing this dynamically as conditions change. Another direction is distributed DTs, where each major component (motor, gearbox, etc.) has its own twin, and they synchronize via a communication protocol (like each running on an edge device in different departments but sharing data). This would mirror how companies are structured and could ease collaboration. A caution is complexity; hence, research into model order reduction for whole-system models is needed. An actionable step is an open-source or reference model of an EV powertrain NVH that can be used as a testbed for such hierarchical twin experiments. If the community has a standard example (like a small EV with a known motor and gearbox model), different teams could try connecting their submodels to form the twin, facilitating best-practice development [10].

7.2. Building Robust Data Pipelines and Feedback Loops

Future Directions—Data Integration 4 (FD–D1): Implement closed-loop digital threads in pilot projects. To overcome broken feedback loops (Gap D1), industry and academia can collaborate on pilot implementations that demonstrate data flowing through the entire lifecycle. For instance, an automotive OEM could choose a particular gearbox program and ensure that at each stage—design, manufacturing, bench testing, field—data is fed to a central twin model and updates are made. Researchers could document how each feedback improved the model or decision. One suggestion is to adopt an iterative model updating framework—e.g., using Bayesian updating, where the twin’s uncertain parameters (like damping or stiffness) are treated probabilistically and updated as new evidence (test data) arrives [25]. A future scenario to aim for is as follows: When a noise issue is found in a durability test, the twin automatically assimilates that result and suggests which parameter to tweak to match it, and that updated parameter is then flagged for design. The tools to do this (like filtering algorithms, etc.) exist, but need adaptation to the NVH context.

Future Directions—Data Integration 4 (FD–D2): Develop and adopt standard ontologies/data models for NVH Twin data. The community should work on defining a common data structure for NVH in DTs. A concrete step could be an extension to ISO 23247 or a new standard focusing on “Digital Twin data for dynamic systems” that includes NVH. It could specify things as follows: how to represent a measured order spectrum in a twin database, how to label operating conditions (load, gear, etc.), and what metadata to store (sensor type, accuracy). In the absence of formal standards, an open-source schema could be proposed by researchers. For example, using JSON or XML to encapsulate NVH results: “component: gearsetA, metric: TE harmonic1 amplitude, simulation: 0.5 arcsec, test: 0.6 arcsec, timestamp: …”. If many start using a similar schema, it could become a de facto standard. Additionally, working with OPC UA companion specifications—OPC UA allows industry-specific “companion specs”; thus, perhaps an NVH Companion Spec could be written defining nodes for “GearMeshFrequency”, “OrderAmplitude”, etc. Another parallel is the FMI standard—there is an FMI for co-simulation; maybe propose an “NVH model exchange” standard or recommended practice. Implementations of these standards in software will be key: if gear testing machine vendors and CAE software can import/export in the same format, the integration gap shrinks. This is a longer-term, yet vital, and more organizational direction.

Future Directions—Data Integration 4 (FD–D3): Closer integration of field data—use of cloud and edge analytics for NVH. For Gap D3 (using field data), one direction is deploying inexpensive NVH data loggers in a sample of vehicles or machines to gather real-world noise/vibration data. Then apply big data analytics to that—cluster analysis to see patterns of NVH issues, correlation with conditions. Already, telematics data is used for maintenance [8]. For NVH, it might involve analyzing sound clips from vehicles.

Future Directions—D4: Address data governance—use secure data sharing and synthetic data. To alleviate security concerns (Gap D4), future work could demonstrate secure pipelines, e.g., using encryption or on-premises DT servers that keep sensitive data in-house while still connecting to cloud compute when needed. Another approach is data synthesis: share less sensitive surrogate data instead of real, or anonymize the data (remove identifying information, normalize absolute values). From a research perspective, developing methods to quantify how much data sharing improves twin accuracy can help justify to management that it is worth solving the IP issues. That quantification can drive internal changes or partnerships.

7.3. Establishing Validation Protocols and Metrics

Future Directions—Validation 1 (FD–V1): Propose a standard validation framework and error metrics. For Gap V1, researchers and standard bodies (like the SAE Vehicle Noise Committee) could work on a recommended practice for validating simulation models of NVH. Such a framework might include: (a) defining key metrics (error in dB for overall levels, frequency error in Hz for peak locations, possibly psychoacoustic metrics differences), (b) defining test conditions needed for validation, and (c) statistical measures. Additionally, since a twin can update, one could include a metric for how quickly it converges after new data—an “adaptability” metric. Creating a benchmark problem (like a specific gear system with known measured NVH) as a community reference would help, similar to how computational fluid dynamics has standard test cases. Perhaps an organization like NAFEMS could start a working group on DT validation—NAFEMS already has some material on simulation credibility.

Future Directions—Validation 2 (FD–V2): Conduct extensive validation studies on diverse cases. To fill the Gap, academic or industry consortia could undertake systematic validation studies. For example, test a gearbox at multiple operating conditions, gather high-quality NVH data, then see how well various modeling approaches can predict those results. Publish the findings. The EU Horizon programs or NSF could fund such comparative studies because they advance general knowledge. The wind industry has conducted slightly analogous studies by comparing different simulation codes against field data for turbine load predictions. Performing this for gear noise—maybe using an established test rig like the FZG gear test rig or an automotive transmission—would highlight where current models succeed or fail. It could reveal, for instance, consistent under-prediction of noise at certain frequencies, guiding which physics to include. Once validated in the lab, an interesting extension is validation in an actual vehicle or machine, to check if the twin holds outside lab conditions (covering Gap V5).

Future Directions—Validation 3 (FD–V3): Incorporate sound-quality evaluation into twin validation. Future NVH twin development should include subjective validation. This can be achieved by synthesizing the twins’ predicted sound and performing listening tests. If the twin is good, listeners should not distinguish between the twin-generated sound and the actual recorded sound of the gearbox within some tolerance. Researchers can use tools like HEAD Acoustics NVH simulator to do “virtual vs. real” comparisons. Additionally, define target metrics for sound quality differences—e.g., the difference in Prominence Ratio between predicted and measured values should be below a certain threshold, since the human detection threshold is known. By integrating psychoacoustic criteria, the twin validation becomes more aligned with the final NVH goals. This is a burgeoning area—literature is sparse; hence, initial studies would break ground. For example, if a twin prediction has all the correct orders but is off in amplitude, does that correlate with subjective annoyance? Possibly not linearly; therefore, incorporating models of auditory perception into twin evaluation would be beneficial.

Future Directions—Validation 4 (FD–V4): Develop trust measures and explainability for hybrid twins. For Gap V4, one approach is to use techniques from explainable AI to make the twins’ ML outputs interpretable. For example, sensitivity analysis to show which input factors most influenced the prediction in a given case (some tools, like SHAP values in ML, can do this). Suppose an engineer sees that “according to the twin, surface roughness had a 5× bigger effect on the predicted noise than torque level in this scenario,” that builds understanding. Another idea is DT health indicators—meta-metrics that indicate when the twin might be off (like if the input goes outside the range it was trained on, raise a flag). Researchers could devise a “consistency check” method—perhaps comparing the ML model’s output to a simpler physics estimate to see if they diverge significantly [25]; applying such methods in the NVH context would ensure the twin is not straying into non-physical predictions without notice. Ultimately, achieving a high level of validation for hybrids might also involve redundancy: having two different models that both predict NVH and cross-validate each other—if they disagree beyond tolerance, the twin knows it is not reliable in that region. This concept of a “self-aware” twin that monitors its own accuracy could be a research frontier.

Future Directions—Validation 5 (FD–V5): Continuous validation in operation and updating validation protocols. Instead of validating once, future twins can be continuously validated as new data comes (like online V&V). Research can explore methods for live validation—e.g., using statistical process control charts on the difference between twin prediction and actual sensor readings during operation, to detect if the twin’s accuracy is degrading. In some sense, the twin itself can be used to detect anomalies by how much it cannot predict (residuals). For example, if the twin consistently fails to predict a certain vibration peak that appears over time, that might mean something physically changed—thus the validation process becomes also a diagnostic tool. This concept is mentioned in condition monitoring literature and can be extended to NVH. Therefore, a future direction—an “adaptive validation” mechanism— is built into the twin.

7.4. Enhancing Application and Sustainability Outcomes

Future Directions—Application 1 (FD–A1): Document and quantify case studies demonstrating twin benefits. To achieve the buy-in (Gap A1), success stories with numbers are needed. Thus, we encourage publishing case studies where a DT actually shortened development or reduced prototypes for an NVH-critical component. For example, an automaker could trial using a DT to replace one round of physical testing on a new EV gearbox and then track differences. If the vehicle meets NVH targets, they can say: Look, we skipped a prototype build, saving X weeks and Y cost. Over the next few years, as early adopters implement NVH twins, they should be encouraged (or required in project grants) to report these outcomes [29].

Future Directions—Application 2 (FD–A2): Explore active control integration (DT-informed control). For Gap A2, research could produce a demonstrator where the DT is in the loop for active noise or vibration control. One idea: an active engine mount whose stiffness can be varied; use the DT’s prediction of upcoming noise to adjust the mount in real time to minimize noise. Research on model predictive control (MPC) using DTs could show how far this can go. This is tricky because NVH is a fast phenomenon and actuators have limits, but it at least offers semi-static adjustments or scenario-based changes. Also, design optimization: using the twin as a test bed to run virtual experiments (design of experiments) to find optimal design or calibration for NVH. This is already performed in CAE, but a connected twin could do it continuously as new data refines the model—i.e., continual optimization. For example, once the twin is validated with prototype data, run an automatic optimization on micro-geometry to further reduce predicted whine, then implement that change in final production. That loop closes the design process faster than traditional build-test-fix. Therefore, developing methodologies for “twin-driven optimization” is a ripe area.

Future Directions—Application 3 (FD–A3): Focus on user interaction and trust-building measures. For Gap A3 (human trust), besides explainability, we need to integrate DTs into the engineer’s daily tools in a friendly way. Management-level trust might come from seeing that the twin has been validated. Another idea is to implement the twin gradually—e.g., running it in parallel with physical tests for a couple of projects and comparing outcomes. If engineers see consistently that “the twin predicted we would be okay and indeed we were,” they will gain confidence. Publishing those internal validations can also help the broader community’s trust. One could also create “when can you trust your twin?” guidelines—e.g., if it is validated within X% in relevant conditions, then it can be used for a Y decision. Having such guidelines from respected bodies or experienced users will help organizations know when it is safe to rely on the twin.

Future Directions—Application 4 (FD–A4): Interdisciplinary training and easy-to-use twin tools. To address Gap A4 (skills gap), educational programs and tooling improvements are needed. Universities might start offering courses combining NVH engineering with data science basics, or DT development for mechanical systems. For example, a GUI where they can drag-and-drop a “machine-learning calibrator” into their model to automatically tune unknown parameters from test data. The easier it is to use, the lower the skill barrier. There is also scope for automation: using AI to assist in model building and updating (some AI can suggest model improvements or find correlations in data that humans miss). An exemplary direction is what one reference mentioned—Lugaresi and Matta (2021) used process mining to rapidly develop DT models [108]. A similar approach could perhaps extract NVH models from test data automatically. These developments will reduce dependency on rare expertise and make twin tech more accessible to average engineers.

Future Directions—Application 5 (FD–A5): Engage with regulatory bodies about simulation approval. For Gap A5 (regulations), the community should proactively work with regulators to find ways to incorporate simulation evidence—maybe initially as a supplement; e.g., allowing one of the required pass-by noise tests to be replaced by a simulation if the simulation process is validated. Additionally, production noise checks can be considered, allowing a DT prediction of noise to waive some end-of-line tests, provided the twin’s predictive accuracy is audited periodically. These are currently long shots; however, with stepwise approaches, it could set a precedent. The benefit to regulators could be consistency. Demonstrating a highly validated twin and sharing it with a standards committee as proof of concept might plant seeds. A near-term action is writing a white paper or standard draft on “Use of DT for Vehicle NVH Development and Certification” that outlines how reliability could be ensured—something automotive industry groups could produce.

In conclusion, the future directions above map closely to each identified gap (summarized in

Table 3. Mapping Gaps to Future Directions (FD).

Gap Category	Key Gaps (From Section 5)	Future Directions (FD) to Address Them
Modeling (M)	M1: Fidelity vs. speed trade-off	FD-M1: Multi-fidelity and reduced-order models; hybrid physics-ML surrogates for fast yet accurate predictions.
	M2: Manufacturing variability	FD-M2: Directly integrate metrology data into models; digital thread from QA to simulation.
	M3: Multi-physics uncoupled	FD-M3: Co-simulate thermal/lube with structural; parametric models linking temperature and NVH; include motor dynamics.
	M4: Lack a system-level model	FD-M4: Hierarchical twins, modular FMU-based assembly of full vehicle NVH; distributed twin architecture.
Data (D)	D1: No closed feedback loop	FD-D1: Pilot closed-loop implementations; Bayesian model updating with each new data input.
	D2: No standard data format	FD-D2: Develop NVH twin data schema/ontology; OPC UA companion specs; push for standardization via SAE/ISO.
	D3: Field data not used	FD-D3: Utilize cloud/edge to collect field NVH data; analytics to feed twin; transfer learning for new scenarios.
	D4: Data sharing concerns	FD-D4: Secure data pipelines (encryption, access control); use synthetic or anonymized data to protect IP; establish data governance policies for twins.
Validation (V)	V1: No std metrics	FD-V1: Propose validation metrics and protocols (error criteria, test matrix) via NAFEMS/SAE; set acceptance criteria.
	V2: Little experimental validation	FD-V2: Conduct comprehensive twin vs. test studies on benchmark gear systems; share results (perhaps in open challenges).
	V3: Lacking sound quality val.	FD-V3: Include psychoacoustic metrics in validation; do listening tests of twin predictions vs. real; calibrate twin to subjective scales.
	V4: ML part not verifiable	FD-V4: Use explainable AI for twin (sensitivity analysis, feature importance); develop self-validation alerts when twin extrapolates; ensure physical constraints in ML.
	V5: Not validated in the field	FD-V5: Perform in situ validation and refine twin; implement online validation (twin tracks and flags when reality deviates).
Application (A)	A1: Benefits not quantified	FD-A1: Publish case studies with ROI and sustainability metrics (prototypes saved, emissions cut); share success stories widely.
	A2: Twin not used in control/opti	FD-A2: Integrate twin with active control (MPC, ANC); use twin for automated design optimization loops; demonstrate predictive control reducing NVH.
	A3: Low trust in twin outputs	FD-A3: Invest in user-friendly interfaces, AR/VR visualization of twin results; gradual twin adoption strategies; involve end-users in twin development to build trust.
	A4: Skills gap and workflow	FD-A4: Cross-train engineers in data science and vice versa; develop high-level twin-building tools requiring less coding; establish best-practice workflows.
	A5: Regulatory acceptance	FD-A5: Engage standards/regulatory bodies to allow simulation evidence; provide guidelines for twin validation in a regulatory context; maybe start with component-level noise regulations.

below). By following these, researchers can systematically close the gaps. Many are already starting in pockets, but a more unified push will accelerate progress.

Implementing these future directions will require close collaboration across disciplines. Mechanical NVH specialists, data scientists, control engineers, and IT experts must work together—reflecting the interdisciplinary spirit of Industry 4.0/5.0 [8]. To illustrate the structure and interdependence of the identified gaps, Figure 7 presents a hierarchical taxonomy of challenges in NVH-oriented DT development. The lower levels capture technical and data-management issues that support all higher functions. Mid-level gaps relate to validation, where missing standards and metrics hinder model credibility. At the top are application and regulatory challenges, involving industrial adoption, trust, and certification readiness.

This taxonomy highlights that progress in higher-level implementation is contingent on resolving the lower-level data and modeling gaps. It thus provides a structured view of how technical, methodological, and organizational barriers interrelate within the NVH digital twin ecosystem.

If these recommendations are pursued, we anticipate that over the next 5–10 years, DTs for gear NVH will move from experimental curiosities to standard tools in the engineer’s toolbox. We will see development cycles with fewer prototypes, NVH issues caught and resolved virtually, and even in-service vehicles benefiting from predictive NVH management. Ultimately, this will lead to quieter, more optimized drivetrains delivered faster and with less environmental impact—realizing the dual goal of technical excellence and sustainability. The relationships between identified research gaps and proposed future directions are summarized in Figure 8.

Beyond methodological and validation considerations, NVH-oriented digital twins also carry significant economic and sustainability implications, which are summarized in Section 7.5.

7.5. Cost–Benefit and Sustainability Aspects

Adopting NVH-focused digital twins (DTs) in drivetrain development delivers both economic and sustainability advantages. These benefits arise from the ability of digital twins to virtualize design and testing, reducing material use, time, and energy while improving decision quality and lifecycle efficiency [109].

Reduction in Physical Prototypes: A major economic gain is the reduction in prototype builds. Traditional drivetrain NVH refinement often involved multiple physical prototypes to identify noise issues and validate countermeasures. With high-fidelity digital twins, many of these steps can be performed virtually, simulating gear whine, vibration modes, and acoustic responses early in development. This not only saves cost but also shortens schedules. Studies show that digital twin use can reduce prototyping costs by up to 40% and cut development time by 20–30%, lowering project durations from 18–24 months to 12–15 months [110,111]. Fewer prototypes directly translate to lower material, labor, and energy use—a clear sustainability gain.

Early NVH Risk Management: Digital twins allow engineers to detect NVH problems early, during concept and design phases, instead of discovering them during late-stage vehicle testing. This “front-loading” of NVH analysis prevents costly redesigns and warranty issues. Simulations can identify resonance issues or tonal noise early, helping avoid expensive post-production fixes. As noted by industry analyses, preventing even one major NVH issue can save millions in recall or warranty costs, while protecting brand reputation [112].

Faster Feedback and Continuous Improvement: When a DT is connected to manufacturing and field operations, it creates rapid feedback loops. Real-time analytics can flag noise-related deviations in machining or assembly, reducing scrap and rework. Rather than relying on reactive troubleshooting, manufacturers can implement immediate, data-driven corrections, improving first-time quality [113]. Lower scrap rates and more stable processes translate directly into energy and cost savings, supporting broader sustainability targets.

Energy and Material Efficiency: Digital twins inherently promote resource efficiency. Fewer physical tests mean less energy spent running dynamometers or producing prototype parts. In addition, optimized NVH design can reduce the need for heavy damping materials, leading to lighter, more efficient drivetrains. Studies show that digital twins in manufacturing can achieve up to 30% energy savings and significant reductions in material waste [114]. NVH optimization may also extend component lifetimes, as predictive monitoring identifies wear before failure, minimizing replacements and environmental impact [115].

Investment vs. Return (ROI): Although DT deployment requires investment in simulation capability, sensing, and data infrastructure, ROI studies consistently report strong gains, typically 10–30% improvements in cost, time, and quality [111]. In automotive settings, initial multi-million-dollar expenditures are often offset by reduced prototype builds and lower warranty costs. The added benefits of faster development and improved quality further reinforce the DT as a strategic long-term asset.

Organizational Preconditions for Success

To realize these benefits, companies need the following:

• Cross-functional collaboration between simulation, testing, and manufacturing teams.
• Skilled personnel in NVH modeling and data analytics.
• Robust IT infrastructure supporting IoT, cloud, and edge computing.
• Strong management support to embed the digital twin into formal development processes.

Firms that meet these preconditions achieve higher returns, shorter cycles, and stronger sustainability outcomes [116].

In conclusion, NVH-focused digital twins provide measurable economic and environmental returns. They reduce prototypes, accelerate development, and minimize material and energy use. Companies like Ford and GM have reported 25–30% design time reduction and improved quality from predictive twin-based processes. When supported by cross-functional teams and a data-driven culture, NVH digital twins evolve from a cost-saving measure into a cornerstone of sustainable, competitive drivetrain development.

7.6. Practical Roadmap for Implementing NVH-Focused Digital Twins

Transitioning from isolated NVH simulations to a fully integrated DT approach requires a structured path. Although many enabling technologies already exist, their effectiveness depends on coordinated progress in modeling, metrology, data management, and organizational workflows. A phased roadmap helps clarify how the industry can evolve from experimental pilots to a routine DT-based NVH development process.

• Short-Term (6–12 months)

Early efforts should focus on a small, low-risk pilot. A simple gear-pair DT—using measured micro-geometry and reduced-order TE models—can run on a cloud platform with minimal infrastructure changes. A basic data link between metrology, CAE, and NVH tests should be established, along with a compact set of validation metrics. These pilots quickly reveal data, workflow, and model-fidelity issues.

• Mid-Term (1–3 years).

Once the pilot phase is complete, the DT can be expanded to drivetrain-level behavior. Hybrid physics–data methods can be introduced, combining multibody models with corrections from EOL or fleet data. Factory-level integration becomes essential, using OPC UA, AAS submodels, or standardized geometry formats to ensure that manufacturing variations flow directly into simulations. Reduced-order models of housing flexibility or acoustic radiation can also be prepared to speed updates. Mid-term work relies on close collaboration between NVH, metrology, and IT teams, enabling validation against full test campaigns rather than isolated lab data.

• Long-Term (3–5 years).

In the final stage, the digital twin becomes a lifecycle tool, supporting design, validation, production, and in-service monitoring. A long-term objective is a continuously updated “living” twin of each gearbox family, incorporating manufacturing deviations, operational loads, and degradation trends. Standardized datasets and benchmark problems for NVH twins should be established to improve transparency and comparability across the industry. At this stage, the twin can support fleet-level predictions of tonal noise risks, early warning of bearing or gear deterioration, and strategic decisions on material usage or prototype reduction. The benefits extend beyond NVH performance, contributing to shorter development cycles, lower scrap rates, and improved sustainability across the drivetrain lifecycle.

• Challenges and Strategies.

Several obstacles may arise along this roadmap. Data silos remain a major barrier, as design, manufacturing, and NVH testing often rely on independent systems with incompatible formats. Early adoption of standardized data interfaces and shared AAS submodels helps mitigate this issue. Organizational boundaries also present challenges: NVH teams, CAE groups, and manufacturing engineers may follow different priorities or validation cultures. Regular joint reviews and integrated workflows help establish common assumptions. Finally, full-order structural–acoustic models can be computationally heavy. Model reduction, surrogate models, and cloud-based execution provide practical ways to balance fidelity with speed.

Taken together, this phased roadmap outlines a realistic path toward deployable NVH-focused digital twins, showing how early pilots can grow into a validated framework that supports both technical performance and long-term sustainability goals. 7.7 Validation of NVH-Oriented Digital Twins

In practice, NVH digital twins are typically validated by comparing their simulated noise and vibration results with measurements from test rigs. Studies have shown that well-constructed multibody and vibro-acoustic twins can reproduce experimental behavior with good accuracy. Similarly, real-time digital twins of gear test benches have shown that simulated vibration and noise signals can closely follow measured time histories. High-fidelity flexible multibody NVH models have also demonstrated substantially better accuracy than simpler approaches when benchmarked against rig tests. In industrial practice, validation increasingly relies on hardware-in-the-loop or co-simulation setups, where CAE models run alongside physical tests, as well as online testing strategies in which the twin is continuously updated with sensor data and compared to live measurements.

Despite these efforts, significant gaps remain. As Horváth and Zelei (2024) note, there is a lack of standardized NVH metrics (even for specific issues like gear surface waviness) and correspondingly a lack of consistent benchmarks for validation [1]. In practice, many twin models are only validated in isolated cases or for limited frequency ranges, and real-time validation frameworks (whereby the twin is automatically correlated with running tests) are still rare. Moreover, simulation conditions often idealize boundary conditions and ignore manufacturing tolerances or operating variability; hence, mismatches between test and model can be large. Horváth and Zelei also emphasize that “digital twins require validation” in order to be trusted [1]. Today, many machine-learning or data-driven NVH predictions remain unverified by physical tests. In short, existing validation is typically performed offline and component-by-component, rather than as a fully integrated, closed-loop process.

Future work is focusing on hybrid workflows and standardization to address these issues. One promising direction is a closed-loop approach that seamlessly links design models and manufacturing data: for example, inline measurement of gear profiles could feed into a twin-driven NVH simulation, whose outputs are checked against end-of-line acoustic tests. Shared NVH benchmark datasets (test-bench recordings and calibrated models) would help the community to compare and improve twin approaches. Efforts to define common protocols are also emerging: industry groups are beginning to adopt ISO/SAE NVH test standards for both laboratory and end-of-line measurements, which could form the basis of standardized twin validation procedures (e.g., using defined drive cycles and sensor locations).

8. Conclusions

This literature review has examined the State-of-the-Art and research gaps in applying DT technology to the NVH optimization of gear systems and drivetrains. We surveyed recent advances in gear dynamic modeling, sensing, and data integration that form the building blocks of a DT, and we identified significant gaps that currently impede the full realization of NVH-focused twins. Key conclusions and takeaways include the following:

• Digital twins hold great promise for gear NVH by enabling continuous model refinement through data feedback and reducing reliance on physical prototypes. Initial implementations (in fault monitoring or targeted studies) demonstrate the potential for predictive noise and vibration management. However, fully operational NVH twins that drive design and control decisions in real time are still rare.
• Several technical gaps must be addressed to build effective NVH twins. High-fidelity gear models need to be made computationally efficient for real-time use (perhaps via reduced-order modeling or hybrid ML approaches). Data from manufacturing (e.g., real tooth surface measurements) and field operations are not yet routinely integrated into simulation models, leading to disconnects between predicted and actual NVH. Crucially, standardized methods to validate and trust digital twin predictions are lacking—currently, no consensus exists on how to quantify a twin’s accuracy or when it can substitute for a physical test.
• A research “gap map” was presented (Table 2), clustering the gaps into Modeling, Data, Validation, and Application categories. This map, supported by ~70 recent references, highlights impacts such as missed feedback loops, insufficient multi-physics representation, lack of NVH twin metrics, and unproven ROI of twins. These gaps mean that today’s gear NVH twins, where they exist, often function as advisory tools rather than authoritative sources for decisions.
• We proposed actionable future directions to bridge each gap (Section 7). These include developing multi-fidelity and physics-informed ML models to balance speed and accuracy, establishing digital threads that feed manufacturing and test data back into models, and creating validation protocols for NVH predictions. We also emphasize human and process factors: building engineer trust through transparency and gradually phasing twins into workflows, as well as quantifying the benefits in terms of cost, time, and sustainability to justify adoption. Table 3 summarizes these recommendations aligned with each gap.
• It is also noteworthy that the NVH digital twin efforts are fully aligned with broader industry trends in sustainability. EV drivetrains, with their heightened sensitivity to gear whine, are a prime application for these techniques—and indeed, many cited studies and projects focus on EV transmission noise. By enabling virtual optimization and reducing prototypes, DTs can help cut development time and material waste, supporting companies’ environmental goals. However, care must be taken to also incorporate considerations of sound quality and human perception into the twin framework to ensure that improvements in simulations translate to real subjective improvements.
• Limitations of this review: While we strove for comprehensive coverage, the field of DTs is rapidly evolving. Some very recent industrial advances may not yet be documented in the literature. We focused on the 2016–2025 period; foundational works before 2015 were included sparingly. The review emphasized tonal gear noise; less has been said about transient or shock noise– those phenomena have their own modeling challenges and could be a subject of future twin research. Additionally, we concentrated on ground vehicle drivetrains; similar principles apply to other gear systems (wind turbines, helicopters), and cross-pollination of ideas between sectors was noted. It should also be noted that some of the identified gaps—such as the absence of standard metrics—reflect the state of the open literature; proprietary industrial methods may exist but are simply not published.

Outlook: Rapid progress is expected in the coming years as Industrial IoT, cloud computing, and AI continue to mature and engineering workflows become increasingly digital. The NVH community is already moving in this direction, with dedicated conference sessions and growing support from major simulation-test platform vendors. We can expect more case studies showing DTs identifying NVH issues earlier or optimizing designs beyond what traditional methods achieve. As such successes accumulate, confidence in DTs will rise, and adoption will accelerate.

Achieving reliable NVH digital twins will require close collaboration across disciplines. Gear designers, NVH engineers, data scientists, and software developers must work together rather than in isolated silos. At the same time, stronger standardization efforts are needed to provide shared frameworks to ensure that companies and research teams do not have to rebuild the same solutions independently.

In closing, DT technology offers a new way to approach vibro-acoustic engineering for gear drives. If the key gaps are resolved, each gearbox may eventually be paired with its own evolving digital counterpart. Such twins can support quieter designs, faster development work, and more informed maintenance decisions. The research directions outlined here provide a practical path forward for both industry and academia. With these advances, the tonal whine of gearboxes—once considered inevitable—may become a problem that can be predicted and controlled.