A Review on Reverse Engineering for Sustainable Metal Manufacturing: From 3D Scans to Simulation-Ready Models

Featured Application

Reverse engineering workflows that convert 3D scan data into simulation-ready models can support sustainable metal manufacturing by enabling accurate geometry reconstruction, faster redesign and process planning, and informed decisions on repair/remanufacturing versus replacement. The reviewed approaches are directly applicable to high-value industrial components (e.g., molds, turbine blades, and marine/energy parts) where reducing scrap, reworking, and lead time yields measurable environmental and economic benefits.

Abstract

Reverse engineering (RE) has been increasingly adopted in metal manufacturing to digitize legacy parts, connect “as-is” geometry to mechanical performance, and enable agile repair and remanufacturing. This review consolidates scan-to-simulation workflows that transform 3D measurement data (optical/laser scanning and X-ray computed tomography) into simulation-ready models for structural assessment and manufacturing decisions, with an explicit focus on sustainability. Key steps are reviewed, from acquisition planning and metrological error sources to point-cloud/mesh processing, CAD/feature reconstruction, and geometry preparation for finite-element analysis (watertightness, defeaturing, meshing strategies, and boundary condition transfer). Special attention is given to uncertainty quantification and the propagation of geometric deviations into stress, stiffness, and fatigue predictions, enabling robust accept/reject and repair/replace choices. Sustainability is addressed through a lightweight reporting framework covering material losses, energy use, rework, and lead time across the scan–model–simulate–manufacture chain, clarifying when digitalization reduces scrap and over-processing. Industrial use cases are discussed for high-value metal components (e.g., molds, turbine blades, and marine/energy parts) where scan-informed simulation supports faster and more reliable decision making. Open challenges are summarized, including benchmark datasets, standardized reporting, automation of feature recognition, and integration with repair process simulation (DED/WAAM) and life-cycle metrics. A checklist is proposed to improve reproducibility and comparability across RE studies.

1. Introduction

The transition toward more sustainable manufacturing is pushing industries to reconsider how metal components are designed, validated, produced, and maintained across their life cycle. In many sectors—such as energy, marine, aerospace, tooling, and heavy machinery—critical parts are often “legacy” assets: complete CAD documentation may be unavailable, modifications may be undocumented, and wear or damage may make nominal models unreliable for structural assessment and process planning. In this context, reverse engineering (RE) is increasingly used not only for geometry replication, but as a manufacturing enabler that reconstructs “as-is” digital models from 3D measurements and links real geometry to mechanical performance and decisions [1,2,3].

The specific contribution of this review is to explicitly treat RE as a decision-grade, simulation-oriented manufacturing workflow rather than a stand-alone digitization task by linking scan-derived geometric choices to mechanical credibility and to measurable manufacturing outcomes. The main “workflow knobs” that govern structural trust (uncertainty, defeaturing/idealization, and boundary-condition transfer) are consolidated, and a lightweight, step-wise KPI inventory is provided to support reproducible sustainability claims at the process-chain level.

The value of RE in metal manufacturing is tied to the creation of a robust digital thread from acquisition to validation: optical/laser scanning and X-ray computed tomography (XCT) can deliver dense geometric information, while processing pipelines convert point clouds/meshes into CAD or directly into analysis-ready representations [4,5,6]. These models support verification, redesign, and process planning, and, crucially, finite element analysis (FEA) for stress, stiffness, and fatigue assessment; however, their industrial usefulness depends on “simulation readiness” (watertightness, feature consistency, meshability, and defensible transfer of boundary conditions and loads).

A persistent barrier to wider adoption is that RE workflows often treat geometry reconstruction and numerical analysis as separate problems, while in practice they are strongly coupled through uncertainty. Measurement noise, occlusions, registration errors, and reconstruction choices (e.g., smoothing and defeaturing) can alter the local curvatures, thicknesses, and filets that govern stress concentrations. As a result, scan-derived geometric deviations can propagate into meaningful differences in predicted mechanical response, especially when accept/reject, repair/replace, or remanufacture/new manufacture decisions rely on simulation outputs [1].

Sustainability adds a second layer of complexity—and opportunity. RE can reduce material waste by avoiding the premature scrapping of high-value parts, minimizing reworking through improved digital alignment and process planning, and enabling targeted repair or remanufacturing instead of full replacement. At the same time, scanning, data processing, and iterative loops carry their own energy and time costs, which are often neglected or reported inconsistently. A sustainability-oriented RE review, therefore, benefits from framing the scan–model–simulate–manufacture chain in measurable terms (scrap, energy proxies, rework, and lead time), particularly for repair-centric workflows [7,8].

Existing reviews have addressed RE methods and tools [2,3] as well as scanning and XCT for industrial and metrological applications [4,5,9]. Other reviews focus on repair and remanufacturing, including additive repair routes where RE is a recurring enabling step [10,11,12]. This review is positioned at the intersection of these streams, with a specific aim: to consolidate scan-to-simulation workflows for metal manufacturing and to frame them through a sustainability lens grounded in measurable manufacturing outcomes.

Compared with prior reviews, the contribution of this work is threefold: (i) RE is treated explicitly as a simulation-oriented manufacturing workflow (not only as geometry reconstruction), (ii) uncertainty-aware links between geometric fidelity and mechanical predictions are highlighted, with emphasis on how reconstruction choices affect stress- and fatigue-critical features, and (iii) a lightweight reporting framework and checklist are proposed to support reproducible sustainability claims at the process-chain level (scrap, rework, energy proxies, and lead time), improving comparability across studies [1,7].

This article is a structured narrative review that synthesizes scan–model–simulate–manufacture workflows for metal components and therefore does not introduce new experimental benchmarks or a single harmonized quantitative dataset. Numerical values reported throughout the paper are used as literature-based, order-of-magnitude anchors to illustrate typical sensitivity trends and should not be interpreted as universal performance guarantees for a given scanner, software toolchain, or part family. The proposed tables, checklists, and KPI inventory are intended to support transparent reporting and decision support; their applicability must be tailored to the component scale, required tolerances, qualification pathway, access constraints, and the availability of traceable reference data. Finally, the sustainability framework is designed as a lightweight, step-resolved screening inventory; decision-grade environmental claims still require a dedicated LCA/LCI with explicit system boundaries, allocation rules, and scenario uncertainty.

The remainder of this review is organized as follows. Section 2 summarizes acquisition technologies and the main sources of geometric uncertainty relevant to metal manufacturing. Section 3 reviews point-cloud/mesh processing and CAD/feature reconstruction pathways, with attention to modeling choices that influence structural predictions. Section 4 focuses on scan-to-simulation practices, including mesh generation, defeaturing strategies, boundary condition transfer, and uncertainty propagation to mechanical responses. Section 5 discusses RE-enabled manufacturing decisions with an emphasis on repair and remanufacturing chains, including their interface with metal repair process routes. Section 6 proposes a concise sustainability and reporting framework for the scan–model–simulate–manufacture chain and provides a practical checklist. Finally, Section 7 outlines open challenges and research directions for robust, automatable, and sustainable RE in metal manufacturing.

To improve readability and provide a compact “map” of the review, Figure 1 summarizes the end-to-end scan–model–simulate–manufacture chain discussed throughout the manuscript. The schematic highlights the verification gates that close the digital thread and the iteration loop that is triggered when requirements are not met. In the remainder of the paper, each stage in Figure 1 is detailed in turn, while Section 6.3 operationalizes the step-wise KPI collection indicated in the figure.

2. Data Acquisition and Geometric Uncertainty in Reverse Engineering for Metal Manufacturing

Reverse engineering (RE) for metal manufacturing can be interpreted as the controlled transformation of a physical component into a decision-grade digital model (CAD/CAE/CAM). While classical RE pipelines describe the sequence from measurement to model reconstruction, industrial RE additionally requires that the acquired geometry is metrologically defensible for its intended task (tolerancing, structural assessment, repair planning) rather than simply “visually correct”. Early works, such as that by Varady and co-workers [13], emphasized that RE quality is governed by a full chain of procedures, encompassing data capture, alignment, reconstruction, and validation, where errors may be introduced and amplified at each stage. This “decision-grade” view is consistent with recent integrated-metrology discussions in advanced manufacturing, which explicitly frame in-line/near-line optical measurement as a process-integrated capability whose value depends on traceability, uncertainty awareness, and task-driven measurement planning [14]. The practical implication is that acquisition cannot be evaluated in isolation: the “right” acquisition strategy is the one that enables stable pass/fail outcomes against downstream decision thresholds (GD and T acceptance, machining allowance, boundary-condition transfer), with uncertainty stated at the feature/task level rather than as a single instrument figure-of-merit.

Accordingly, the acquisition step is discussed together with the dominant sources of geometric uncertainty that control model credibility and, ultimately, simulation readiness. To provide a compact overview, Figure 2 summarizes an “acquisition-to-model uncertainty map” that highlights where geometric uncertainty is typically created and how it propagates toward simulation-ready models. The map is intended as a reading guide for Section 2: each uncertainty source is linked to the corresponding control/reporting items (

Table 1. Acquisition modalities vs. uncertainty sources vs. manufacturing fit for metal manufacturing reverse engineering (RE).

Acquisition Modality	Typical Use in Metal Manufacturing	Strengths	Dominant Geometric Uncertainty Sources	Best Choice (Rule of Thumb)
Structured light/laser triangulation	Fast digitization; external geometry; fixtures; worn surfaces	High point density; short cycle time; portable/near-line options	Surface reflectivity/specular highlights; surface condition; calibration; angle-of-incidence effects; occlusions	External surfaces + fast turnaround; when internal features are not required
Handheld optical scanning	Maintenance; in situ digitization; remanufacturing; field inspection	Portability; accessibility; rapid capture in constrained environments	Operator dependency; inconsistent coverage; registration drift; surface-condition sensitivity	Field/maintenance scenarios; fast “as-is” capture with local validation artifacts
Terrestrial laser scanning (TLS)/photogrammetry	Large assets; assemblies; plant/ship structures; retrofit and fit-up	Scalability; large volumes; efficient global shape capture	Range/angle effects; environmental stability; multi-station registration error; occlusions	Large structures; combine with local high-accuracy scans for critical interfaces
X-ray computed tomography (XCT)	Internal channels; hidden features; complex AM parts; volumetric inspection	Internal + external geometry; non-destructive volumetric insight	Beam hardening/scatter; voxel size; reconstruction artifacts; surface determination/segmentation bias	When internal geometry governs function/repair decisions; when destructive inspection is not acceptable

and

Table 2. Minimal reporting checklist (metrology + uncertainty) to support reproducibility and decision-grade use for GD and T and CAE. (GD and T: geometric dimensioning and tolerancing).

Step	Minimum Information to Report (Manufacturing-Oriented)	Why It Matters
Surface condition	Material; finish (machined/shot-blasted/polished); coatings/sprays used	Surface–sensor interaction drives bias/noise (especially on metals)
Acquisition setup	Modality; key settings; standoff; fixturing; coverage strategy	Supports reproducibility and comparability across studies and setups
View planning	Number of views/stations; overlap rationale; occlusion handling	Coverage and registration stability determine local fidelity
Registration	Method (targets/features/best-fit); acceptance criteria; drift checks	Controls compounded alignment error and interface fidelity
Reconstruction	Filtering; decimation; meshing; surfacing; defeaturing rules	Reconstruction choices can shift stress-critical features
XCT specifics (if used)	Voltage/energy class; voxel size; artifact mitigation; segmentation method	XCT “surface determination” is a major uncertainty driver
Uncertainty statement	Task-specific uncertainty or proxy (feature-level if possible)	Turns scan data into decision-grade geometry for tolerancing/CAE
Validation	Comparison vs. reference (CMM, gauge, artifact) or repeatability checks	Prevents overconfidence and supports simulation readiness claims

) and to the downstream steps where it most strongly affects simulation credibility (Section 3 and Section 4).

In addition, Table 1 compares acquisition modalities in terms of manufacturing fit and dominant uncertainty drivers, while Table 2 provides a minimal reporting checklist to improve the reproducibility and comparability of scan-based RE studies. Beyond supporting reporting, these elements are intended to make acquisition choices “decision-facing”: they indicate which uncertainty class is most likely to dominate a given use-case and therefore which controls (surface conditioning, view planning, datum strategy, XCT surface determination settings) should be treated as qualification levers rather than implementation details.

To connect the acquisition families discussed in Section 2 to practical industrial choices, Figure 3 provides a compact selection map based on geometry type (external vs. internal), component scale, and typical hybrid setups. The schematic also clarifies how multi-sensor acquisitions are registered/fused in a common reference frame (targets, features, datums) before generating CAD- or analysis-ready outputs. The implication is that “hybrid” acquisition should be treated as a planned architecture (global reference + local functional interfaces), because registration stability and datum definition often dominate the credibility of downstream tolerancing and scan-to-simulation boundary transfer.

2.1. Main Sources of Geometric Uncertainty in Scan-Based RE

Across metal manufacturing applications, the dominant contributors to geometric uncertainty can be grouped into seven recurring classes: (i) surface–sensor interaction (reflectivity, roughness, coatings), (ii) calibration and system stability, (iii) view planning and occlusions, (iv) multi-view registration and drift, (v) filtering/meshing/surfacing choices, (vi) XCT artifacts and surface determination/segmentation, and (vii) task-specific interpretation of uncertainty for tolerancing and simulation readiness. As an order-of-magnitude benchmark, controlled close-range optical scanning has demonstrated errors below 5 μm with expanded uncertainties below 10 μm on calibrated miniature step gauges [15], whereas a metrological XCT example reported a calibrated voxel size of 80.005 μm ± 0.001 μm (1σ) and a propagated standard uncertainty contribution of ±1.01 μm on a 55 mm bi-directional length [16]; at larger scales, terrestrial laser scanning typically operates in the sub-millimeter to millimeter range for range error, depending on instrument class and conditions [17]. These contributors are intentionally reflected in the structure of Section 2 and serve as a bridge to the scan-to-CAD and scan-to-simulation discussions in Section 3 and Section 4 (see Figure 2). From an implementation standpoint, the key point is that “dominant uncertainty” shifts with scale and task: at a small scale, it is often surface–sensor interaction and processing; at a large scale, it is frequently registration network stability and environmental control; and for XCT, it is commonly surface determination and artifact mitigation.

A recent, practically relevant trend across several of these classes is the use of “metrology digital twins” (and related error/uncertainty maps) to turn repeated artifact-based scans into spatially resolved corrections and uncertainty estimates that can be propagated to GD and T features and downstream decisions [18]. In a structured-light case study, such a digital twin produced a volumetric error-vector field and uncertainty mapping, halving the mean bias (≈−10 to −5 μm) and reducing error dispersion by ~27% after correction [18]. The implication is that uncertainty maps can serve as a direct enabler for automation: they allow downstream steps (GD and T checks, scan-to-CAD, BC transfer) to use confidence-aware thresholds rather than relying on conservative global margins.

2.2. Optical 3D Scanning of Metals: Structured Light and Laser Triangulation

Structured-light and laser triangulation scanners are widely adopted on the shop floor because they enable the rapid acquisition of dense surface geometry. Artifact-based benchmarking shows that achievable model precision is strongly feature-dependent even under controlled conditions: using a structured-light scanner (GOM ATOS Q), Jacobs et al. [19] reported point spacing of 0.04–0.15 mm and a diameter repeatability on a 15.9 mm circular feature with a 2.5 μm range (0.0025 mm) and 0.5 μm standard deviation (0.0005 mm), while some circles’ center-to-center distances involving smaller features reached ranges of up to 40 μm (0.04 mm).

However, on metallic parts, their performance is often dominated by surface–light interaction rather than by nominal device specifications. Javaid et al. [20] provided an industrial overview of 3D scanning and emphasized that effective deployment requires selecting a sensing principle based on the inspection task while accounting for shop-floor constraints such as line-of-sight/access, fixturing, operator-related variability, and cycle time. In the context of digital manufacturing, Catalucci et al. [21] highlighted that uncertainty and robustness remain key obstacles when optical metrology is pushed closer to production, where surface condition variability and environmental effects cannot be avoided.

For metals, the most critical issue is that reflectivity and local surface texture can cause saturation, multipath reflections, and biased point placement near edges and steep slopes. Surface conditioning (e.g., matte coatings) is therefore often used to improve data completeness; however, coatings can introduce thickness bias and should be treated and reported as a measurable process variable. Recent evidence also clarifies the order of magnitude of “surface-driven” effects: in a metrology-oriented comparison of laser triangulation configurations, probing form error was in the 0.022–0.044 mm range and combined standard uncertainty was about 0.011–0.022 mm, while the raw point-cloud dispersion (σ) varied roughly from 10 μm up to ~78 μm depending on surface condition and sensor setup—values that can easily dominate over nominal point spacing when scanning metals [22]. Accordingly, coating selection and application thickness should be reported alongside scanning settings, and ultra-thin coatings have been shown to reduce coverage loss by ≈30% compared with thicker alternatives, highlighting a measurable completeness–bias trade-off [23,24].

In this direction, Bonin et al. [25] proposed an artifact-based characterization approach for handheld laser scanners, demonstrating how local and global deviations can be quantified in a practical manner rather than inferred from single-number accuracy claims. Such characterization is particularly relevant for manufacturing-oriented RE because it supports defensible acceptance criteria and traceable comparisons across setups. Beyond “instrument accuracy”, recent manufacturing-facing studies show that the data-conditioning chain itself can be a first-order uncertainty driver: for additively manufactured metal parts, Giganto et al. [26] showed that σ-based point-cloud filtering primarily affects geometrical outcomes: using 2σ–3σ filters yielded geometric results comparable to contact reference values, whereas a more restrictive 1σ filter distorted the geometry and could even artificially produce form errors below the reference.

Similarly, comparative industrial studies continue to refine the question of “how to benchmark scanners” under realistic constraints; for instance, Poboża et al. compared optical laser scanning systems and reported accuracy/suitability trade-offs that are directly relevant when selecting devices and validation artifacts for production-oriented workflows [27].

A second practical trend is the shift from operator-driven scanning to automation. Recent studies on robotized optical scanning report improved repeatability by controlling standoff distance and viewing angles and by reducing operator variability—an aspect that becomes decisive when scan data are intended for downstream CAD/CAE automation [28]. In the same direction, model-based/robot-guided viewpoint planning has been explicitly proposed to make optical 3D digitization more repeatable and coverage-driven—i.e., less dependent on operator heuristics [29]. For simulation-oriented RE, the practical implication is that repeatability and consistent coverage are often more valuable than marginal gains in point density.

2.3. Large-Scale Metal Assets: Terrestrial Laser Scanning, Photogrammetry, and View Planning

For large metal structures and assemblies (tooling, plant assets, shipbuilding components), terrestrial laser scanning (TLS) and photogrammetry offer scalability, but uncertainty becomes increasingly coupled to coverage completeness and registration stability across many stations. Modern TLS instruments operate over tens to hundreds of meters, with range errors typically spanning less than a millimeter to several millimeters in range, with range noise often on the order of a few hundred micrometers; angular uncertainties are commonly in the tens-of-arc-seconds range, and environmental variations can become non-negligible in large-volume setups. This motivates traceable, procedure-based performance evaluation rather than relying on nominal datasheet specifications alone. A dedicated review of TLS evaluation discusses key error sources, performance testing, and traceability aspects that are directly relevant when RE outputs are used for retrofit design, alignment, or fit-up decisions [17]. Broader overviews of large-scale optical dimensional metrology also remain useful to position TLS and photogrammetry within a unified manufacturing metrology landscape, especially when hybrid setups are used (TLS for global shape + local high-accuracy scanning on functional interfaces) [30,31].

Recent industrial metrology work has also revisited “photogrammetric scale” as a practical accuracy limiter: Trombini et al. [32] evaluated scaling photogrammetry approaches for industrial metrology applications, providing evidence-based guidance on when (and how) photogrammetry can be made competitive for large assets. Their quantitative benchmarking reports relative (expanded) uncertainty values that decrease with an increasing measurand, reaching about 0.2% at meso-scale; as a practical translation, this corresponds to ~2 mm on a 1 m characteristic length, which is often the accuracy band relevant for large-asset retrofit or fit-up use-cases.

In these scenarios, view planning is a first-order design variable because it affects occlusions, incidence angles, and the number of required stations, which in turn determine both the achievable uncertainty and the acquisition effort. Peuzin-Jubert et al. [33] surveyed the view planning problem for reverse engineering and automated applications, clarifying how viewpoint selection governs completeness and the risk of registration drift in multi-scan pipelines. A more recent review focused on autonomous view planning for 3D scanning of unknown/partially known geometries and provides an updated synthesis of sensing/coverage objectives and planning strategies that map well onto inspection and retrofit contexts [34]. For manufacturing-oriented RE, the practical implication is that viewpoint network conditioning (coverage, overlap, incidence angle distribution, and station geometry) should be treated as an uncertainty control lever and reported explicitly, because registration stability can become the dominant contributor once workflows span many stations. Complementary industrial approaches also frame view planning as a “digital-twin optimization” problem for photogrammetry/camera networks, explicitly optimizing camera placement to improve reconstruction precision under production constraints [35].

2.4. Industrial X-Ray Computed Tomography (XCT): Internal Geometry and “Surface Determination”

Industrial XCT is uniquely valuable in metal manufacturing when internal geometry must be captured (cooling channels, internal cavities, lattice cores, hidden damage/defects) or when external access is constrained. Foundational overviews establish XCT as a dimensional metrology tool while clarifying its dominant limitations and practical trade-offs. The CIRP keynote by Kruth et al. [4] is a cornerstone reference that establishes XCT as a dimensional metrology tool while outlining its dominant limitations—beam hardening, scatter, voxel size effects, reconstruction artifacts, and the need for traceability and uncertainty evaluation. De Chiffre et al. [5] provided a complementary overview of industrial CT applications and constraints, explicitly addressing the practical trade-offs among resolution, speed, and usability for production contexts. Villarraga-Gómez et al. [6] further summarized the transition from imaging to dimensional metrology, emphasizing how accuracy drivers and surface extraction strategies control the final dimensional reliability. As a manufacturing-facing quantitative anchor, when scale calibration, artifact mitigation, and surface extraction are controlled, deviations versus traceable tactile/CMM references can fall in the low-tens-of-micrometers range for small-to-medium metal artifacts; conversely, for large or dense parts requiring increased field-of-view (larger voxels), “CAD-ready” surface fidelity commonly shifts toward ~0.1 mm (and beyond) if not compensated through calibration and robust surface determination [4,5,6]. Consistently, international benchmark exercises/inter-comparisons show that best-practice setups can achieve sub-voxel length accuracy (≈0.1 voxel) on calibrated artifacts, underscoring that voxel size provides a baseline but not the sole determinant of final model accuracy [36].

For metal RE, one recurring critical step is surface determination (often implemented through segmentation/thresholding decisions). Because different extraction strategies can yield different geometries from the same voxel dataset, surface determination directly affects wall thickness and curvature details that are mechanically relevant, and thus the credibility of simulation-ready models. Methodological studies explicitly addressing XCT surface determination (segmentation and surface extraction) are important anchors when XCT-based reverse engineering is used to justify simulation-ready models, because different extraction strategies can lead to materially different geometries and local feature fidelity [37,38,39,40]. A practical rule-of-thumb emerging across studies is that thresholding/surface-determination choices can shift the extracted boundary by a fraction of a voxel up to ~one voxel, mapping directly into wall-thickness/curvature bias and non-negligible volume differences in thin walls, lattice struts, and small radii—features that often drive structural/thermal simulations [37,41,42,43,44].

Within this methodological line, recent work continues to propose and benchmark surface determination algorithms tailored to CT metrology needs; for example, Yang et al. [45] presented a surface determination algorithm for CT metrology using a simplex/tetrahedral mesh strategy, and Pirillo et al. [46] proposed an improved surface determination technique explicitly discussed in the context of industrial CT metrology.

In addition, manufacturing-led discussions on “design for XCT” and measurement planning provide a useful bridge between metrology and engineering arguments, because they highlight how fixturing, orientation, exposure settings, and part geometry control artifacts and, ultimately, the credibility of reconstructed surfaces [47]. For instance, orientation (penetration length) measurably affects dimensional errors: favorable setups can reduce length deviations to below 10 μm under metrology-grade conditions, while unfavorable orientations increase artifact-driven bias and uncertainty [48]. Surface condition is another key driver: for rough metal surfaces (cast/AM), CT–tactile offsets can become comparable to the roughness amplitude, and roughness should therefore be treated explicitly in the uncertainty budget when CT-derived surfaces are used for tolerancing or simulation boundary conditions [49].

When AM-related internal complexity is involved, XCT becomes even more central as a link between geometry and performance. Thompson et al. [9] reviewed XCT for additive manufacturing, identifying barriers that generalize well to metal manufacturing RE: resolution limits, surface measurement challenges, speed/cost constraints, and the difficulty of translating voxel data into dimensional statements suitable for engineering decisions. Where RE objectives explicitly include volumetric measurands (e.g., cavity/porosity volume fraction, void morphology, internal defect sizing), calibrated objects and reference artifacts have been used to benchmark volumetric accuracy and software-to-software variability, enabling traceable volume statements rather than purely qualitative defect maps [50]. Finally, a recent review specifically addresses how machine learning is being used across the industrial XCT pipeline (from reconstruction/denoising to defect analysis and metrology-relevant tasks), providing an up-to-date entry point for “ML-assisted XCT” in manufacturing practice [51].

2.5. Representative Industrial Systems (Illustrative Examples)

To connect the methodological discussion to common industrial practice, representative off-the-shelf systems can be mapped to the main scanning families considered in this review. For structured-light/optical 3D metrology, industrial scanners such as ATOS Q (ZEISS, Oberkochen, Germany) are widely used for high-resolution full-field surface acquisition in inspection and RE contexts. For handheld laser scanning (often preferred for portability and shop-floor flexibility), metrology-grade devices such as HandySCAN 3D (Creaform, Lévis, QC, Canada) represent common industrial solutions. For large components and environments where long stand-off distances are required, terrestrial laser scanning platforms such as FARO Focus (FARO Technologies, Lake Mary, FL, USA) and Leica RTC360 (Leica Geosystems, Heerbrugg, Switzerland) are representative examples. Finally, for internal geometry and defect inspection, industrial X-ray CT systems used in metrology and NDT include platforms such as ZEISS METROTOM (ZEISS, Oberkochen, Germany), Phoenix V|tome|x (Waygate Technologies, Hürth, Germany), Nikon XT H series (Nikon Metrology Europe NV, Leuven, Belgium), and Werth TomoScope (Werth Messtechnik, Giessen, Germany). These examples are intended solely to illustrate typical equipment categories and do not represent an exhaustive list; comparable systems exist across multiple vendors.

2.6. Registration and Alignment: Turning Multiple Scans into Coherent Geometry

Because registration/alignment is a frequent failure mode in scan-based workflows, Figure 4 summarizes a minimal, implementation-oriented alignment pipeline and distinguishes target/datum-based from targetless strategies. The intent is to explicitly state the minimum acceptance checks (e.g., target residuals/ICP RMS/overlap) and the resulting reliability mapping that should accompany the aligned dataset before downstream GD and T or CAE use. The schematic also makes explicit the minimum quality checks required before aligned data are used for tolerancing, scan-to-CAD reconstruction, or simulation.

Multi-view acquisition is typical for both optical scanning and TLS, so registration is often a dominant uncertainty contributor. Classical rigid registration frameworks remain useful for describing the underlying logic of this process, even when commercial software provides “black-box” implementations. Besl and McKay [52] introduced the iterative closest point (ICP) framework, which still underpins many alignment pipelines in RE. Chen and Medioni [53] later discussed object modeling via multi-view range images, providing a complementary perspective on multi-view integration that is frequently referenced when explaining range-data fusion logic. Recent surveys further consolidate rigid registration methods with updated evaluation practices and datasets, supporting more defensible choices of registration settings in manufacturing RE [54,55].

For manufacturing use, the key implication is that the chosen strategy (target-/datum-based, feature-based, best-fit) determines whether residual deviations concentrate on functional interfaces or are distributed globally, directly affecting tolerancing checks and boundary-condition transfer for CAE; accordingly, Table 2 includes explicit reporting items for registration strategy and acceptance criteria.

2.7. Task-Specific Uncertainty: Making Scan Data Usable for Tolerancing and Simulation Readiness

To support decision making, scan-based geometry should ideally be accompanied by a task-specific uncertainty statement (even if simplified). In GPS terms, when scan-derived geometry is used to support acceptance decisions (fit-up, sealing, machining allowance, or simulation boundary conditions), it is good practice to report decision rules consistent with ISO 14253-1 and uncertainty statements consistent with the GUM (ISO/IEC Guide 98-3). Wilhelm et al. [56] discussed the concept of task-specific uncertainty in coordinate metrology, emphasizing that uncertainty depends on the measurement task and strategy rather than on the instrument alone. Although originating in CMM contexts, the same logic transfers directly to scan-based RE: uncertainty must be expressed at the feature/task level (e.g., fatigue hotspot, sealing interface, press-fit diameter), not as a single global “scanner accuracy” value.

Recent literature has increasingly explored virtual and digital-twin-based uncertainty estimation for scanning systems, aiming to reduce the need for repeated physical trials. Vlaeyen et al. [57], for example, proposed a digital twin approach for the uncertainty estimation of optical scanning systems integrated with coordinate measurement, illustrating how virtual metrology can support task-specific uncertainty statements without disrupting production workflows. A closely aligned, task-specific formulation for laser scanning explicitly builds the uncertainty budget via a digital twin and Monte Carlo simulations within the ISO GPS/verification logic, making it particularly relevant for tolerancing-oriented RE [58]. Complementarily, open-access work demonstrates how digital-twin-enabled error/uncertainty maps can be generated for structured-light scanning and propagated through correction and GD and T steps, providing a practical pathway from scanner characterization to decision-grade point clouds [18].

For manufacturing-centric RE, this uncertainty-aware perspective is a prerequisite for repeatability, traceability, and automation, and it directly strengthens scan-to-simulation credibility.

Therefore, acquisition and uncertainty reporting should be structured around the downstream decision gate (GD and T acceptance, machining allowance, BC transfer, or QoI-based simulation), because the same scan can be sufficient for one task yet non-credible for another.

For clarity, the acronyms used in the figures/tables are as follows: BC = boundary conditions; QoI = quantity of interest; KPI = key performance indicator; GD and T = geometric dimensioning and tolerancing; TLS = terrestrial laser scanning; XCT = X-ray computed tomography; ICP = iterative closest point.

2.8. Brief Note on Sustainability

While sustainability is addressed in detail in Section 6, it is worth stating one principle here: acquisition choices can be sustainability-relevant mainly through their impact on reworking, scrap prevention, and iteration count. In practice, the “best” acquisition modality is not necessarily the least energy-intensive one, but the one that delivers adequate confidence for the intended manufacturing decision with the least total iteration and downstream waste. This principle motivates the reporting checklist in Table 2 and the uncertainty map in Figure 2.

3. From Scan Data to Simulation-Ready Digital Models: Processing, Reconstruction, and Structural Sensitivity

Building on the acquisition stage summarized in Section 2 (see Table 1 for sensing options and Table 2 for dominant uncertainty sources), this section addresses the post-acquisition steps that most strongly control whether scan-derived geometry becomes analysis-credible rather than merely analysis-eligible: point-cloud/mesh processing, CAD/feature reconstruction, and the preparation of simulation-ready models. The key message is that “cleaning” and “reconstruction” are not neutral operations: each modeling choice can selectively damp, shift, or even create geometric details that later drive stress concentrations, stiffness predictions, and fatigue-relevant local fields. The practical implication is that processing parameters should be treated as decision-facing levers: a setting that improves visual smoothness or meshing robustness can still bias stress-critical radii, notch roots, or contact interfaces if it is not tied to a target quantity of interest (QoI) and an error-control rationale.

From a practical tooling perspective, the processing steps discussed in this section are typically supported by widely adopted software ecosystems for point-cloud inspection/processing and scan-to-CAD reconstruction. Representative commercial examples include PolyWorks (InnovMetric, Québec City, QC, Canada), Siemens NX/Simcenter toolchain components (Siemens Digital Industries Software, Plano, TX, USA), and CATIA/3DEXPERIENCE–SIMULIA environments (Dassault Systèmes, Vélizy-Villacoublay, France). In parallel, open-source alternatives are also used in industrial and academic contexts (open-source community, International, online), especially for point-cloud cleaning/visual inspection and mesh handling, and can be integrated in the early stages of scan-to-model workflows depending on qualification requirements and traceability constraints. In practice, software naming is secondary to traceability: the critical requirement is that key processing operations (filtering, healing, decimation, feature fitting, and simplification thresholds) are recorded and reproducible, because they define the effective “as-used” geometry that drives downstream CAD/CAE decisions.

3.1. Registration and Metrology-Aware Point Sets

Most industrial reverse-engineering (RE) workflows start from multiple views (or multiple datasets) that must be brought into a common coordinate frame. A comprehensive reference for registration strategies (rigid and non-rigid, pairwise and multi-view) is the survey by Tam et al. [59]. Lyu et al. [54] provide a consolidated view of rigid registration methods together with common evaluation protocols and datasets, while Marek and Chmelař [55] offer a complementary taxonomy that highlights practical limitations and failure modes across registration families.

In manufacturing-oriented RE, the practical point is that registration uncertainty is rarely uniform: it depends on surface finish, incidence angles, occlusions, fixturing repeatability, and the presence of “weak” geometric constraints (e.g., near-planar regions). For rigid mechanical parts, a classical and still widely used baseline is the Iterative Closest Point (ICP) framework by Besl and McKay [52].

However, even “good” registration can blur small filets, shift hole axes, or bias wall thickness when alignment is underconstrained—effects that become critical when the downstream goal is structural simulation rather than pure visualization.

A robust practice (especially for metals) is to preserve a raw dataset (unfiltered) alongside a processed dataset, and to document registration settings (targets vs. targetless alignment, datum choices, number of views, overlap percentage). This supports traceability when simulation/experiment discrepancies later need to be explained. This traceability mindset is consistent with recent discussions of learning-based point-cloud workflows in industrial inspection and measurement, where reproducibility and uncertainty-aware pipelines are treated as design requirements rather than afterthoughts [60].

3.2. From Points to Mesh: Normals, Watertightness, and Repair

When the target is a surface model, the workflow typically proceeds through normal estimation, triangulation, and then mesh repair. For reconstructing surfaces from oriented points, a seminal method is Poisson surface reconstruction by Kazhdan et al. [61]. Huang et al. [62] offer survey and benchmark surface-reconstruction strategies from point clouds, providing a useful reference to motivate reconstruction choices under the noise, sparsity, and missing-data conditions typical of manufacturing scans.

Although developed in the geometry-processing community, Poisson-type reconstructions are attractive in RE because they are comparatively resilient to noise; the trade-off is that the “smoothness prior” can attenuate sharp edges or small radii unless constraints (or post-processing) preserve them.

Once a mesh exists, it is often simplified to reduce size and improve downstream usability. A classic and still influential approach is quadric-error-based simplification by Garland and Heckbert [63]. For structural use, simplification is a modeling decision rather than a purely computational step, because even small edge rounding or loss of narrow grooves can shift local stresses in notches, filets, and thin features.

For FE-driven workflows, watertightness and topological consistency are often mandatory. A representative “mesh-healing” approach that explicitly targets imperfect models (gaps, overlaps, T-joints, small holes) to enable FE model generation is the work by Chong et al. [64]. Complementary to this line of work, Wen et al. [65] introduce a feature-preserving mesh repair strategy that is particularly relevant when small radii and sharp transitions must be retained for structural analyses.

In practice, healing and decimation rules (gap-closure thresholds, hole-filling policies, self-intersection handling) should be set relative to the minimum structurally meaningful feature size for the specific component. A compact checklist of these post-processing choices—registration, denoising, reconstruction, healing, and decimation—and their structural implications is provided in

Table 3. Geometry-processing choices and their typical structural implications, linking each point → mesh → CAD operation to the model changes it introduces and to the stress-/fatigue-relevant consequences that may affect simulation credibility.

Step/Choice (Point → Mesh → CAD)	What Changes in the Model	Why It Matters for Structural Predictions	Typical Control Parameter	Ref.
Registration strategy (datum vs. targetless; multi-view)	Global pose; small systematic shifts; axis/feature alignment	Misalignment can bias thickness/axes → stiffness and stress errors; affects BC transfer	ICP settings; overlap; datum constraints; target layout	[52]
Outlier removal/denoising	Removes sparse extremes and high-frequency variation	Can erase sharp edges/small grooves mistaken as noise; can shift filets	Neighborhood size; statistical thresholds; filter type
Normal estimation	Orientation field for reconstruction and smoothing behavior	Drives edge fidelity and reconstruction stability; affects local curvature	kNN radius; robust normal filters; smoothing strength
Surface reconstruction (e.g., Poisson)	Mesh topology and smoothness; hole closure tendencies	Smoothness prior can attenuate sharp/stress-critical features and small radii	Octree depth; regularization; screening/point weights	[61]
Hole filling/gap closure	Watertightness; local geometry inference	Filling may remove vents/holes that are load paths or stress raisers; may alter stiffness	Max hole size; curvature-based fill rules; patch constraints
Mesh healing	Repairs topology and geometry defects (gaps, overlaps, self-intersections)	Enables meshing but can “invent” geometry if tolerances are loose; may distort interfaces	Healing tolerance; intersection resolution rules; sliver-face removal	[64]
Mesh simplification/decimation	Reduces triangles; redistributes error; smooths small detail	Changes radii/filets → alters stress concentration and fatigue hotspots	Error metric bounds; feature-preservation flags	[63]
Simplification (defeaturing and idealization)	Suppresses small features; changes topology	Reduces solve cost but introduces analysis error that should be bounded/estimated	Feature-size threshold; defeaturing rules; error control metric	[66]

to support the transition from mesh processing to CAD reconstruction.

3.3. Segmentation and Surface Extraction as “Stress-Relevant” Decisions

A central step in CAD reconstruction is the partitioning of point/mesh data into patches that can be assigned a geometric meaning (planes, cylinders, blends, freeform regions). A useful manufacturing-oriented example is the region-growing strategy for dense, noisy meshes by Vieira and Shimada [67].

Segmentation is not only a geometric classification problem; it effectively defines where edges, blend zones, and feature boundaries will lie in the reconstructed CAD, thereby governing curvature continuity and local radii—parameters that often dominate stress concentration.

For metal components (machined, cast, or additively manufactured with post-processing), this is especially critical around small filets/chamfers, hole entrances/counterbores, transitions between planar faces and freeform regions, and thin ribs/bosses.

Recent contributions on segmentation and feature extraction in mechanical/CAD-like geometries include Romanengo et al. [68] on recognizing geometric primitives in point clouds of mechanical CAD objects, Fugacci et al. [69] on reconstructing and preserving feature curves during point-cloud processing, and Zhang et al. on the interactive reverse engineering of CAD models via sequential primitive construction [70].

3.4. Surface Fitting and NURBS Reconstruction: Tolerances That Matter

For freeform or mixed-geometry parts, CAD-ready surfaces are frequently represented as NURBS patches fitted to the measured data. An example of a NURBS fitting strategy explicitly positioned for reverse engineering is the approach by Dan et al. [71].

The crucial modeling variable is the fitting tolerance (and any smoothing/regularization): tighter tolerances preserve local waviness (potentially including measurement noise), while looser tolerances generate visually “clean” surfaces that may underrepresent small curvature changes and filet transitions.

For simulation-ready models, a helpful rule is to relate fitting tolerance to (i) expected acquisition uncertainty (Section 2, Table 2), and (ii) the smallest radius or thickness that influences the targeted failure mode (e.g., fatigue notch sensitivity vs. global stiffness). A recent feature-preserving reference explicitly focused on reconstructing CAD-like surfaces with sharp features from noisy/non-uniform point clouds is Cai et al. (“FACE”) [72], which provides a feature-preserving perspective on the tolerance–sharpness–noise trade-off that often governs whether local radii are retained or inadvertently smoothed.

3.5. Feature and CAD Reconstruction Pathways: From Meshes to Parametric Solids

A classical framing of RE for CAD reconstruction is given by Várady et al. [13], who describe the pipeline from measured data to usable geometric models and discuss the role of segmentation, surface fitting, and design intent.

Engineering CAD reconstruction additionally requires recovering topology (faces/edges/loops) and, when possible, feature semantics (holes, pockets, blends) so that the output is not only geometrically plausible but also usable downstream. A relevant algorithmic contribution in this direction is Benkő et al. [73], who address methodologies for reconstructing boundary-representation models in RE contexts.

When RE must be “manufacturing-usable” (editable and constraint-aware), knowledge-based strategies become important. Fisher [74] discusses the use of knowledge in RE problem solving, highlighting that purely geometric reconstruction may not capture the functional structure needed for downstream tasks. From a workflow-driven perspective, Bénière et al. [75] provide a comprehensive mesh-to-CAD process, including data preparation and CAD-entity construction steps that reflect practical reconstruction pipelines. From an industrial perspective, an early but still frequently cited overview of RE in manufacturing is provided by Kruth et al. [76]. Alongside these established pipelines, the recent literature has increasingly targeted data-driven reconstruction of CAD representations that remain editable and constraint-aware. A dedicated survey on deep learning for CAD reconstruction [77] helps position where AI-based methods can (and cannot yet) replace feature-aware pipelines, while deep-network-driven reconstruction of editable parametric CAD models [78] directly connects RE outputs to manufacturing-relevant CAD. The major reconstruction pathways discussed in this section are summarized in

Table 4. CAD reconstruction pathways for manufacturing RE and simulation readiness, contrasting the main reconstruction routes in terms of output type, practical strengths, and typical limitations that can affect simulation readiness (e.g., feature fidelity, continuity, and editability).

Pathway	Output	Strengths	Limits (Simulation View)	Ref.
Surface-driven (segmentation → NURBS patches)	NURBS surfaces/stitched B-Rep	Good for freeform; controllable fitting tolerance	Patch boundaries may not coincide with functional features; continuity issues can affect stress	[67,71]
Feature-driven (primitives + blends + topology)	Feature-aware CAD	Preserves design intent; easier edits and downstream CAM	Harder on worn/as-built geometry; depends on robust segmentation and tolerance models	[13,73]
Knowledge-based reconstruction	Parametric CAD with constraints	Better functional consistency; supports standard feature libraries	Requires domain knowledge; not purely data-driven; may be difficult to automate	[74]
Mesh-to-CAD workflow (industrial pipeline)	CAD-ready geometry from mesh	Practical end-to-end procedures; fits industrial toolchains	Quality depends strongly on preprocessing and mesh repair; risk of over-smoothing	[75,76]
Simulation-oriented simplification	Analysis model (idealized CAD/mesh)	Reduces meshing and solve cost; easier parametric studies	Must quantify or bound analysis error; defeaturing can remove hotspots	[66,79,80]

, which describes contrasting outputs, strengths, and typical structural-analysis implications, and can be found at the end of Section 3.5.

3.6. Preparing “Simulation-Ready” Models: Defeaturing, Idealization, and Error Control

Even when a CAD model is available, simulation commonly requires defeaturing and idealization (removing small blends, suppressing holes, simplifying filets, and dimensional reduction). A widely used entry point is the survey by Thakur et al. [79] on CAD model simplification techniques for physics-based simulation.

In this review, “model simplification” is used as an umbrella term for operations that make a geometry analysis-ready. Two distinct actions are considered: (i) defeaturing, i.e., removing/suppressing explicit local features (e.g., holes, small blends/filets) that mainly affect topology and stress raisers; and (ii) idealization, i.e., replacing portions of the model with simplified representations (e.g., mid-surface/shell, symmetry, contact idealization, dimensional reduction) to reduce complexity while preserving the targeted physical behavior. Accordingly, “simplification” is used for the general concept, while “defeaturing” or “idealization” denotes the specific operation.

For simulation-driven RE, the critical gap is that defeaturing is not “free”: it introduces analysis error that should be estimated or at least controlled. Li et al. [66] provide a quantitative framework for estimating the effects of removing negative features on engineering analysis.

Similarly, Tang et al. [80] propose an index to evaluate defeaturing-induced impact on FE model analysis.

Finally, simulation readiness is often blocked by CAD quality issues (gaps, invalid trims, inconsistent tolerances) that prevent stable meshing or cause solver failures. A focused survey on quality assurance and testing tools for CAD models is given by González-Lluch et al. [81].

At this stage, the RE output is typically analysis-eligible (e.g., repaired mesh or CAD/B-Rep) but not yet analysis-credible, because meshing robustness, boundary-condition transfer, and uncertainty propagation still govern the reliability of the predicted mechanical response. Section 4, therefore, shifts to scan-to-simulation practices, detailing how meshing, simplification choices, BC definition, and uncertainty quantification affect credibility; the analysis-aware treatment of simplification (including error control tied to specific QoIs) is discussed there. Therefore, Section 4 treats scan-to-simulation as a QoI-driven credibility problem, where meshing and boundary-condition transfer are qualified against decision thresholds rather than geometry-only metrics.

4. Scan-to-Simulation Practices: Meshing, Defeaturing, Boundary Conditions, and Uncertainty Propagation

Scan-to-simulation (S2S) workflows aim to translate scan-derived geometry into mechanical predictions that are credible enough for design, verification, and manufacturing decisions. In practice, S2S is rarely a linear pipeline: it is an iterative loop in which geometric representations are refined, mesh generation is re-tuned, boundary conditions (BCs) are revisited, and uncertainty is progressively constrained. The practical implication is that “simulation readiness” is not achieved by geometry quality alone: credibility is reached only when the chosen representation, mesh robustness, and BC assumptions remain stable under reasonable variations in processing and alignment settings, and when residual uncertainty can be expressed in terms of QoI-relevant bounds.

Table 5. Scan-to-simulation workflow for reverse-engineered metal components, summarizing the main pipeline steps from geometry ingestion to validation, the key modeling choices at each step, and the typical mechanics-driven failure modes that compromise simulation credibility.

Scan-to-Simulation Step	Typical Inputs	Key Methods/Choices	Main Failure Modes (Mechanics View)	Ref.
(1) Geometry ingestion	Point cloud/mesh/B-Rep; optional XCT volume	Choose analysis representation: direct mesh, reconstructed CAD, or immersed/implicit domain	Loss of stress-critical radii; inconsistent topology; non-physical smoothing	[62,75]
(2) FE mesh generation	CAD/mesh (possibly defective), feature tags, target element size	Tetra meshing; hex-dominant; adaptive sizing near filets/holes; defect-tolerant meshing	Slivers/inverted elements; mesh leakage through gaps; unresolved thickness	[82]
(3) Simplification (defeaturing and idealization)	CAD or reconstructed B-Rep; feature list; analysis objectives	Suppress small holes/filets; mid-surface extraction; idealize contacts	Removal of fatigue hotspots; biased stiffness; altered contact stresses	[83,84]
(4) Material assignment	Nominal alloy data; local hardness/HT state; AM/repair zones if any	Piecewise material zones; anisotropy when needed; calibration via model updating	Wrong local stiffness/strength in repair/HT zones; over/under-predicted safety factors	[85]
(5) Boundary condition (BC) transfer	Test/fixture definition; measured displacements/forces; scanned contact interfaces	Map BCs from CAD/test to FE mesh; infer constraints; contact modeling	Non-physical constraints; load path errors; unrealistic stress patterns	[80]
(6) Validation and iteration	Measured strains, deflections, or DIC; inspection metrics	Compare predicted vs. measured response; update geometry/BCs; uncertainty statements	Overfitting; missing model-form error; non-reproducible pipelines	[85]

provides a compact “S2S checklist” mapping the typical steps of a scan-to-simulation workflow to the most common mechanics-driven failure modes and representative references, while

Table 6. Boundary-condition transfer and uncertainty propagation in scan-to-simulation, linking common S2S problems (BC mapping/inference and uncertainty sources) to practical approaches and to the minimum reporting items needed to make results reproducible and decision-grade.

Problem	Practical Approaches	Pros/Limitations	What to Report (Minimum)	Ref.
BC transfer between non-matching models	Nearest-neighbor/projection; RBF mapping; mortar methods; constraint equations	Fast but may distort near edges; robust methods need careful parameterization	Mapping method; smoothing/regularization; validation check vs. measurements
BC inference from measurements	Inverse methods; FEM model updating; regularized optimization	Can identify hidden constraints; sensitive to noise/model error	Measurement type (DIC/strain/displacement); cost function; regularization	[85]
Geometric uncertainty propagation	Perturbation/sensitivity; Monte Carlo with shape variants; surrogate models	MC is robust but costly; sensitivity may miss nonlinearity; surrogates need training	Uncertainty source model; number of samples; convergence criteria
Mesh/repair uncertainty	Compare multiple reconstruction/repair parameter sets; feature-level tolerances	Captures processing variability; needs systematic experiment design	Parameter ranges; acceptance thresholds; effect on hotspot stress	[65]
As-built surface-driven fatigue hotspots	Use scanned topography as geometry input; local refinement	Captures real roughness; needs careful filtering/scale separation	Scan resolution; filtering cutoffs; notch definition; fatigue metric	[86]

focuses on the two most error-sensitive aspects in industrial settings: boundary-condition handling and uncertainty propagation.

At the meshing and CAE stage, representative commercial environments commonly adopted in industry include ANSYS (Ansys, USA, Canonsburg), Altair HyperWorks/HyperMesh (Altair, Troy, NY, USA), and SIMULIA Abaqus (Dassault Systèmes, Vélizy-Villacoublay, France). For specific communities and use cases, open-source solvers and preprocessors are also used as complementary options (open-source community, International, online), particularly in research-oriented validation pipelines and benchmarking. In practice, the solver ecosystem is less critical than the reporting of the modeling decisions it implements: meshing controls, defeaturing/idealization thresholds, BC transfer assumptions, and the uncertainty-propagation strategy define the effective credibility envelope of the prediction and therefore must be traceable.

4.1. Geometry-to-Analysis Representations: CAD-Based, Mesh-Based, and “Implicit/Immersed” Options

A first decision in scan-to-simulation is what representation to analyze. Figure 5 should be read as three practical routes: CAD-based meshing, direct use of scan-derived meshes, or implicit/immersed representations when CAD reconstruction is not viable. Many workflows reconstruct a CAD/B-Rep and then mesh it (Section 3), but alternative strategies analyze scan-derived meshes directly or use implicit/immersed representations when CAD is unavailable or too costly. Recent benchmarking work on surface reconstruction helps justify these choices under realistic scan imperfections (noise, sparsity, missing regions), and is therefore a useful anchor when considering “why a representation is appropriate” [62]. In manufacturing RE, the “right” choice depends on which features dominate the response: global stiffness might tolerate stronger idealization, while fatigue hotspots typically demand local fidelity at radii, edges, and surface transitions.

4.2. FE Mesh Generation on Scan-Derived or Reconstructed Geometries (Including Defective CAD)

Mesh generation is often the first place where a seemingly plausible reconstruction fails mechanically: sliver elements, inverted tets, non-manifold surfaces, and tiny gaps can trigger solver divergence or yield non-physical stiffness. This is amplified in reverse engineering because reconstructed CAD often contains small defects (gaps, overlaps, self-intersections) introduced during surfacing, trimming, or healing (Section 3).

A highly relevant and recent contribution by Yang et al. [82] focuses on surface mesh generation for industrial CAD models with defects, which explicitly addresses robustness to imperfect geometry. For scan-to-simulation, the practical implication is that “mesh success” is not merely a geometric objective; it is a robustness objective—meshing should be stable across reasonable variations in reconstruction parameters (filters, smoothing, repair tolerances).

Mesh control should be tied to structural sensitivity. Local refinement is typically required at stress concentrators (holes, filets, sharp transitions), and when the scan is used to represent as-built radii rather than nominal ones. When meshing follows defect repair, feature-preserving strategies can be valuable to avoid inadvertently flattening sharp transitions. For example, Wen et al. propose a feature-preserving mesh repair approach based on restricted power diagrams [65], which aligns well with the need to retain sharp features that dominate local stress fields.

4.3. Defeaturing and Idealization for Simulation: Controlling Analysis Error Rather than “Removing Detail”

Defeaturing is unavoidable in industrial simulation: it reduces meshing complexity and eliminates irrelevant detail. However, in scan-to-simulation, the risk is particularly high, because scans naturally contain high-frequency detail (including roughness, waviness, and artifacts) and because small geometric features may be functionally relevant even if they are “small”.

Analysis-aware defeaturing should therefore be framed as an error-controlled operation linked to a target quantity of interest (QoI). In line with the terminology adopted in Section 3.6, this analysis-aware approach is discussed here under the umbrella of simplification, while referring explicitly to defeaturing when the operation consists of suppressing/removing local geometric features. Buffa et al. [83] discuss analysis-aware defeaturing with a posteriori estimation concepts that support selecting simplification levels while monitoring their impact on the mechanical response. In addition, Hinz et al. [84] address domain simplification with controlled accuracy, reinforcing the same S2S logic: simplification is acceptable when its effect on the chosen QoI (e.g., peak stress at a notch, contact pressure, stiffness) can be bounded or verified.

A practical rule for S2S is to base defeaturing thresholds on (i) the acquisition/processing uncertainty (Section 2, Table 2; Section 3, Table 3), and (ii) the failure mode: fatigue and contact problems are typically far more sensitive to small radii and surface transitions than global stiffness or low-frequency modal analysis. Accordingly, Table 3 helps identify which geometric operations are most likely to bias stress-critical features, while Table 6 summarizes the minimum reporting items to make defeaturing choices and their response impact traceable.

4.4. Boundary-Condition Transfer: From Fixtures/Tests/CAD to the FE Model

Boundary conditions are often the dominant “hidden variable” in scan-to-simulation. Even with a high-quality geometry and mesh, incorrect constraint modeling can yield stress patterns that look plausible but are physically wrong. In reverse engineering, BC definition is complicated by incomplete documentation of fixtures, unknown contacts, and non-ideal supports.

Practical BC transfer issues typically include mapping loads/constraints from CAD or test descriptions onto non-matching FE meshes, representing contact interfaces using scanned “as-is” geometry, and inferring effective constraints from measured response when supports and fixture stiffness are uncertain.

A helpful mechanics-oriented concept here is that BC transfer should be treated as a modeling step with validation, not as a one-time setup. For example, Tang et al. propose an index to evaluate the impact of defeaturing on FE analysis [80]; while not a “BC paper” per se, it illustrates the style of thinking that is needed for BC transfer too: it quantifies impact on quantities of interest, rather than assuming neutrality. In practice, Table 6 provides a reporting checklist for BC mapping (method, smoothing/regularization, and validation against measurements).

4.5. Uncertainty Propagation: From Geometric Variability to Mechanical Response Variability

Once scan-based geometry is used for simulation, uncertainty must be considered at two levels:

• Geometric uncertainty from acquisition and processing (surface interaction, registration drift, filtering/smoothing, segmentation thresholds, CT surface determination—Section 2; processing biases—Section 3).
• Simulation uncertainty conditioned on geometry (mesh sensitivity, BC uncertainty, material variability, and solver/model-form uncertainty).

A key point for manufacturing-oriented S2S is that geometry uncertainty is not necessarily small: small changes in local radii or notch-like features can change peak stress and fatigue indicators substantially. This is particularly visible when as-built topography is explicitly used. For instance, a recent applied study integrates 3D-scanned surface topography into an FE-based fatigue workflow for WAAM steel and demonstrates the centrality of surface-driven hotspots [86]. This type of contribution is valuable for scan-to-simulation reviews because it makes the link explicit: scan fidelity → local stress fields → fatigue metrics.

Practically, uncertainty propagation is typically performed via the following:

• Perturbation/sensitivity analysis (fast, but may miss nonlinearity);
• Monte Carlo with geometric variants (robust, but costly);
• Surrogate models (efficient after training, but require careful validation).

The minimum reporting items for geometric uncertainty propagation (uncertainty model, number of samples, convergence criteria, and how hotspots are tracked across variants) are summarized in Table 6.

4.6. Model Updating and Response-Driven Refinement (Closing the Loop)

In industrial settings, S2S workflows often rely on feedback from measured responses (displacements, strains, DIC) to refine uncertain parameters such as BC stiffness, contact conditions, or material zones—especially for repaired or additively restored parts. A recent review on finite element model updating (FEMU) for material calibration provides an up-to-date entry point to this broader “response-driven refinement” logic and its uncertainty implications [85]. While FEMU is not exclusively a scan-to-simulation method, it becomes highly relevant when scan-based geometry is used as the geometric backbone, and the remaining uncertainty concentrates in BCs and material parameters. Therefore, decision-grade S2S requires reporting not only a nominal result, but a QoI-focused credibility statement that links geometry provenance, BC assumptions, and uncertainty propagation to the accept/reject or repair-planning decision threshold.

5. RE-Enabled Manufacturing Decisions: Repair and Remanufacturing Chains for Metals

Reverse engineering (RE) becomes manufacturing-relevant when its outputs are not treated as “a model” but as decision-grade evidence that supports (i) repair vs. replace choices, (ii) repair-route selection, and (iii) qualification/return-to-service gates. In metal manufacturing, this is particularly visible in high-value components (e.g., blades, molds, tooling, heavy assets) where the economic and technical rationale of remanufacturing depends on a reliable digital thread from damage capture → repair planning → process execution → verification. A comprehensive recent synthesis of additive manufacturing-based remanufacturing for repair/restoration is provided by Kanishka and Acherjee [10], which also clarifies why RE is repeatedly required at multiple stages (initial assessment, repair volume definition, and post-repair verification).

To keep this section compact and easily citable, the key “decision gates” and the minimum digital evidence typically required at each gate are summarized in

Table 7. RE-enabled repair/remanufacturing decision gates and the digital evidence typically required, showing where RE outputs enable specific go/no-go decisions and which metrics and uncertainty sources most strongly affect each gate.

Decision Gate (Where RE Enables a Choice)	Digital Inputs (RE + Data)	Key Engineering Criterion	Typical Metrics	Dominant Uncertainty Sources	Ref.
Repair vs. replace/scrap	Initial inspection + 3D scan or CMM; optional XCT for internal damage; damage map vs. nominal CAD	Is the remaining life/strength recoverable within allowable limits?	Damage volume; min remaining wall thickness; crack length; allowable stock for machining; tolerance band for functional interfaces	Scan completeness (occlusion); registration drift; defect detectability; surface extraction (XCT) if used	[10,87]
Repair-volume definition (additive vs. subtractive patch)	Alignment of damaged geometry to nominal; extraction of deviation field; segmentation of repair region	Repair patch must remove damaged material while minimizing heat-affected risk and preserving datum features	Repair volume; patch boundary curvature continuity; datum distance-to-damage; minimum radius retained	Registration bias near weak constraints; smoothing/meshing choices; patch boundary sensitivity	[88,89]
Process-route selection (LC/DED/WAAM/LPBF vs. hybrid ASM)	Damage class + access; required accuracy; deposition accessibility; material compatibility; optional process simulation	Route must meet geometry + property targets with acceptable risk (cracking/porosity/distortion)	Achievable bead size; heat input class; post-machining allowance; deposition rate; minimum feature size	Thermal history variability; material dilution; geometry-to-toolpath fidelity; monitoring coverage	[90,91,92]
Toolpath planning and collision/pose feasibility	Reconstructed surface + local normals; reachability and collision checks; scan-derived thickness maps for finishing	Feasible paths that control bead overlap and avoid overbuild on constrained regions	Local deposition angle; standoff variation; overlap ratio; reachable orientation window; expected overbuild	Local normal noise; surface reconstruction bias; robot/cell calibration	[89,90]
Parameter-window selection and stability control	Geometry-informed heat accumulation risk; in situ monitoring signals (melt pool/height/IR)	Stable deposition without lack-of-fusion, cracks, or excessive distortion	Linear/areal energy density; layer height error; melt pool size; interpass temperature	Emissivity variation; sensor line-of-sight; powder/wire feed fluctuations	[91,93,94]
Post-processing and qualification (machining, HIP/HT, re-inspection)	As-built scan/CT for verification; simulation/inspection plan; mechanical testing strategy	Return-to-service acceptance on geometry + microstructure/property + defect limits	Final dimensional deviation; porosity/crack thresholds; hardness/UTS; residual stress indicators	Metrology of freeform repaired regions; CT surface determination; sampling representativeness	[92,95,96]
Network/make-or-buy (central hub vs. distributed repair)	Digital thread portability (scan + CAD + process plan); accuracy target; logistics/capacity data	Cost–quality–lead-time trade-off for where repair is executed	Accuracy threshold; transportation cost; capital amortization; turnaround time	Provider capability variance; data interoperability; calibration equivalence	[97,98]

. The major metal repair routes and their interface requirements with RE outputs (geometric inputs, planning constraints, and verification expectations) are contrasted in

Table 8. Repair/remanufacturing process routes for metals and the interface requirements with RE outputs (geometry, planning, and verification), contrasting the main repair families by what RE must deliver (repair volume, datums, access, allowance) and how verification is typically closed to manage route-specific risks.

Repair Route	Typical Metal Components	What RE Must Provide	Process-Model/Monitoring Interface	Post-Processing and Verification	Key Risks	Ref.
Laser cladding/powder DED (LP-DED/DED-LB/M)	Turbine blades, molds, rails, high-value housings; local wear and edge rebuild	Accurate repair volume + boundary; surface normals; access map; machining allowance envelope	Thermo-mechanical simulation for distortion/residual stress; online QC for melt pool/height/defects	Finish machining; heat treatment/HIP when needed; re-scan + targeted NDT/CT	Cracking (Ni alloys), porosity, dilution, distortion; overbuild in constrained regions	[93,94,96]
Wire-laser DED (W-LDED)	Near-net rebuild on steel/Ti; repair where powder handling is undesirable	Continuous pathable boundaries; stable standoff/orientation planning; bead stacking strategy	Height/melt-pool control; temperature monitoring; sensitivity to travel direction and thermal history	Machining for tolerance; re-scan for surface form; metallography for dilution/bond quality	Geometry drift from heat accumulation; bead waviness; sensitivity to path planning	[91,93]
Wire Arc Additive Manufacturing for remanufacturing (WAAR/WAAM-based)	Large structures and heavy components; thick-section rebuild; shipbuilding/mining parts	Global registration across large parts; collision-free posture planning; adaptive slicing/toolpath	Heat accumulation control; bead geometry prediction; in-process sensing, where possible	Subtractive cleanup/machining; re-scan of functional interfaces; mechanical testing as needed	High heat input → distortion; coarse microstructure; limited resolution for small features	[90]
LPBF-based repair/rebuild (PBF)	Small-to-medium high-value parts; complex local rebuild; superalloy feature restoration	High-fidelity repaired region definition; support strategy constraints; interface preparation geometry	Build simulation for residual stress/distortion; defect mitigation via post-treatment planning	HIP/HT; re-CT (porosity) + re-scan (geometry); microstructure at interface	Porosity/cracking; interface metallurgical compatibility; anisotropy and HAZ	[92]
Hybrid additive–subtractive remanufacturing (ASM/in-cell scan–build–machine)	Precision repair with tight tolerances; turbine blades/blisks; complex tooling	Datum strategy; additive/subtractive patch segmentation; scan-to-machine coordinate consistency	Cost/sequence optimization; verification scans between steps; simulation for removal/addition order	Integrated finishing; intermediate inspection; final metrology vs. tolerance requirements	Error stack-up across steps; fixturing repeatability; data handoff between software modules	[97,98,99]

.

5.1. From Geometry to Decisions: The Repair/Remanufacturing “Gate Model”

A practical way to frame RE-enabled remanufacturing is as a sequence of gates: (1) repairability screening, (2) repair volume definition, (3) process–route selection, (4) toolpath/process planning, (5) stabilization/monitoring, and (6) verification and qualification. The important point is that RE is not “upstream only”: scan-based evidence often returns after manufacturing steps for intermediate and final verification, closing the loop between planning and execution. Table 7 consolidates these gates with the typical metrics (damage volume, minimum remaining thickness, machining allowance, accuracy thresholds) and where geometric uncertainty most strongly impacts the decision.

Recent work on automated inspection/assessment illustrates why this gate framing matters at scale: for example, Zhang et al. [87] propose a vision-based approach for defect detection on heavily rusted parts as a front-end enabler of remanufacturing triage. While not a geometry-reconstruction paper per se, it highlights a recurring industrial reality: the remanufacturing decision begins with damage detectability and classification, before any CAD/CAE reconstruction is attempted.

5.2. Damage Assessment and Repair-Volume Definition: Extracting the “Patch” That Manufacturing Can Execute

The transition from “damaged geometry” to a manufacturable plan typically requires (i) alignment to a nominal reference and (ii) extraction of a repair patch (additive and/or subtractive). A representative hybrid remanufacturing algorithm that explicitly integrates measurement, reconstruction, repair patch extraction, and decision logic is given by Zheng et al. [88]. Although earlier than the latest wave, it is still useful here because it makes the repair-volume concept concrete and connects it directly to downstream manufacturing steps.

More recent work pushes this toward automation for complex freeform components. Friebe et al. [89], for example, propose an automated toolpath planning method for turbine blade repair explicitly driven by 3D scan data, including a robust way to link a degraded blade to an appropriate target geometry. In the same spirit—but at a broader workflow level—Chen et al. [95] discuss laser metal additive remanufacturing for turbine blade repair and emphasize the integration of RE-derived repair volumes with CAD/CAM toolpath generation in repair chains.

One implication for a manufacturing-focused review is the following: the repair-volume definition should be presented as a manufacturing interface problem—the patch boundary must be geometrically stable (registration-robust), process-feasible (accessible orientations), and inspection-feasible (verifiable tolerances). This is why Table 7 explicitly lists the repair-volume definition as its own decision gate.

5.3. Repair Route Selection for Metals: What Each Process Needs from RE

Once a repair patch is defined, the next decision is selecting a route that can deliver both geometry and properties with acceptable risk. For metals, the dominant families include laser cladding/powder DED, wire–laser DED, WAAM-based remanufacturing, LPBF-based repair approaches, and hybrid additive–subtractive remanufacturing (ASM). Because these routes differ in deposition resolution, heat input, accessibility constraints, and post-processing burden, they impose different requirements on RE outputs and on how verification is conducted. Table 8 contrasts these routes in terms of what RE must provide (repair volume, normals, access maps, datum strategy, machining allowance envelope) and how verification is typically closed (re-scan, CT/NDT, mechanical testing strategies).

A focused recent review on wire–laser DED (W-LDED) provides a useful route-level synthesis of parameter control, stability, and monitoring considerations—topics that directly determine how stringent the RE-to-toolpath geometry must be [91]. For WAAM-based remanufacturing, Du et al. [90] explicitly review data-processing aspects (including reverse reconstruction and planning steps) from a remanufacturing perspective, making it a strong anchor for “RE meets heavy repair” scenarios. For powder–laser DED, a recent review in the Journal of Laser Applications focuses on LP-DED in the context of defect/microstructure/mechanical behavior in functionally graded builds—relevant when repair is used for graded or tailored surfaces [94].

LPBF-based repair appears less common than DED for many field repairs, but it is increasingly used for high-value parts under controlled conditions. Wang et al. present a methodology for repairing damaged nickel-based turbine blades via LPBF followed by HIP, with detailed interface characterization and mechanical testing—useful as a modern “repair route exemplar” where RE-defined repair geometry must remain compatible with build constraints and post-treatment plans [92].

5.4. Planning and Sequencing in Hybrid Additive–Subtractive Remanufacturing

In industrial remanufacturing, hybrid additive–subtractive chains are attractive because they combine material restoration (additive) with tolerance recovery (subtractive finishing). A cost-driven process planning framework for hybrid additive–subtractive remanufacturing is presented by Zheng et al. [98], explicitly positioning process planning as an optimization problem rather than an ad hoc choice. At the system level, Li et al. [99] demonstrate a 6-axis hybrid additive–subtractive manufacturing process (robot-based), which remains a useful reference for discussing coordinate consistency and error stack-up across additive and subtractive steps.

One manufacturing-centric interpretation is the following: in hybrid chains, RE is not only “reverse reconstruction” but also datum governance—the scan-to-machine coordinate relationship and the re-inspection strategy between steps often dominate final tolerance capability. This logic is captured in the “error stack-up across steps” row of Table 8.

5.5. Verification, Monitoring, and Qualification Loops: Closing the Digital Thread

For RE-enabled repair to be credible, it must close with verification that is aligned with the intended use (tolerancing, structural performance, return-to-service). Recent work increasingly treats process stability and online quality control as part of this chain rather than as optional instrumentation. Ye et al. [93] review online quality control for laser DED, emphasizing multi-modal monitoring and the move toward adaptive control—highly relevant when repaired geometry must match a scan-derived target under variable thermal conditions. On the modeling side, Zhou et al. [96] present a multiscale multiphysics simulation of powder-based DED for surface repair (including stress/deformation implications), illustrating how process simulation can be used to pre-validate parameter windows for a given repair geometry.

Recent applied work already demonstrates end-to-end, industrially realistic “inspect–plan–deposit–post-process–reinspect” loops that operationalize the digital thread. For example, Al-Musaibeli and Ahmad [100] propose a layer-by-layer rebuild strategy where depth-sensor measurements of worn surfaces are reconstructed and directly converted into robot-assisted laser cladding toolpaths, enabling a practical scan → reconstruct → path-plan → deposit workflow. Imam et al. [101] demonstrate an autonomous robotic laser cladding repair cell that couples vision-based damage detection with laser triangulation for 3D damage quantification and calibration, specifically addressing the inspection/localization bottleneck for automated repair. In high-value industrial repair, integrated implementations explicitly close the loop from acquisition to post-processing and reinspection: Wang et al. [102] report an integrated automatic system in which a structured-light 3D scanner generates point clouds, damaged regions are identified by comparison against the nominal model, robot programs are generated for cladding, and a dedicated reprocessing step is included—explicitly closing the loop at the implementation level. A related dual-robot implementation further details hand–eye calibration and coordinate transfer from scanner to cladding robot, validating a complete inspection → cladding workflow experimentally [103]. Finally, closed-loop process control is exemplified by inter-layer 3D scanning used to update deposition based on a geometry error criterion (demonstrated on 316L and Inconel 625) [104].

From a standards perspective, scan-to-model and scan-to-repair workflows are typically anchored to three pillars: (i) performance verification of the selected measuring system (e.g., ISO 10360-13 for optical 3D CMS; ISO 17123-9 for terrestrial laser scanners; and VDI/VDE 2634-2/-3 for optical 3D measuring systems), (ii) decision rules and uncertainty statements consistent with ISO 14253-1 and the GUM (ISO/IEC Guide 98-3), including substitution-based uncertainty evaluation where applicable (ISO 15530-3), and (iii) interoperable data exchange for point clouds and derived models (e.g., ASTM E2807/E57). For XCT-based dimensional metrology, ISO 10360-11 has been discussed for CT-specific acceptance testing but has not been consistently available as a finalized International Standard; therefore, industrial practice commonly cites ASME B89.4.23 and VDI/VDE 2630 for performance evaluation and test methodologies. Finally, because repair itself must be qualified, scan-to-repair studies should explicitly link the metrology chain to the applicable process-qualification standards, such as ISO 15614-1 for repair/build-up welding and ISO/ASTM 52920 (qualification principles) and ISO/ASTM 52904 (metal PBF process control) when AM-based repair routes are used.

To complement the methodological discussion with a concrete industrial deployment, Figure 6 shows an in situ optical scanning workstation installed in a shipyard environment within the SHIPLEARNING project. The example illustrates typical shop-floor constraints (large parts, limited access, non-ideal surfaces) and the acquisition-to-model workflow supported by real-time visualization.

Practical rule for scan-enabled repair chains: verification is most effective when it mirrors the decision gates—i.e., re-scan after key steps (after build-up, after rough machining, after final finish), using acceptance criteria that reflect what the repaired part must do (fit, seal, carry load), not just global surface deviation.

5.6. Brief Note: Where the Remanufacturing Decision Is Executed (Networks and “Repair-As-a-Service”)

Finally, the repair/remanufacturing decision is increasingly shaped by where capabilities reside (central hubs vs. distributed partners). A recent quantitative operations-oriented model explicitly evaluates the economic feasibility of cloud-enabled hybrid additive–subtractive repair networks and shows how accuracy thresholds influence routing and hub selection [97]. For a manufacturing-focused RE review, this is useful to motivate interoperability and traceable digital artifacts: scan data, reconstructed geometry, and process plans must be portable across facilities if remanufacturing is to scale. Therefore, industrial credibility in RE-enabled remanufacturing depends less on any single algorithm and more on whether the full gate sequence (Table 7) can be executed with traceable evidence, quantified uncertainty, and repeatable verification loops that remain valid when the workflow is distributed across sites and partners.

6. Sustainability and Reporting Framework for the Scan–Model–Simulate–Manufacture Chain

Reverse engineering (RE) becomes sustainability-relevant in manufacturing when it enables lifetime extension, scrap avoidance, and iteration reduction across the scan–model–simulate–manufacture workflow rather than merely digitizing geometry. In circular decision contexts (repair/remanufacture vs. replacement), sustainability conclusions are often driven less by “how green the scanner is” and more by whether the RE chain delivers a right-first-time manufacturing decision with bounded performance risk and minimized material/energy overhead. Comparative studies on metal additive repair/refurbishment show that material/feedstock and electricity mix frequently dominate environmental hotspots, while process-chain modeling clarifies where improvements are structurally actionable [105,106].

6.1. Compact Framing: Sustainability as a Two-Scenario Question (Repair/Remanufacture vs. Replacement)

A concise and easily citable way to report sustainability for RE-enabled metal manufacturing is to state the problem as a two-scenario comparison: (i) a reference scenario (replacement/new manufacturing) and (ii) an RE-enabled scenario (repair/remanufacture enabled by scan–model–simulate). This immediately forces clarity on the functional unit (e.g., “one component meeting specification over X operating hours”) and on the assumed lifetime extension. Because lifetime modeling can materially affect outcomes, the lifetime assumption should be stated explicitly and, where possible, treated as a scenario variable [107].

From a manufacturing viewpoint, this framing aligns naturally with repair process chains, where the goal is not “a model” but a released part that meets acceptance criteria with minimal additional resource use.

6.2. Where Sustainability “Moves the Needle” in RE Workflows

Across metal manufacturing case studies, sustainability leverage tends to cluster in three places:

• Avoiding rework loops and scrap through metrologically defensible inputs: poorly reported acquisition/reconstruction choices often trigger re-scans, repeated cleanup, or late discovery of mismatches—each iteration adding time, energy, and potentially scrap.
• Selecting the repair/manufacturing route and controlling material yield: in repair chains, deposited/removed mass and feedstock production typically outweigh “digital” energy costs, and outcomes are sensitive to powder demand and electricity mix [105,106].
• Using simulation to prevent over-processing and over-repair: S2S is sustainability-relevant when it reduces conservative repair volumes, heat treatments, or machining allowances while still meeting mechanical requirements; integrated LCA/LCC perspectives are increasingly used to formalize this trade space.

6.3. A Minimal Inventory and KPI Set for the Scan–Model–Simulate–Manufacture Chain

To keep reporting lightweight (and reproducible), the recommendation is to collect a minimal set of inventory items and KPIs per chain step, prioritizing what typically drives conclusions (material, energy, yields, iterations, lifetime). Table 9 synthesizes (i) the main sustainability lever at each step, (ii) what to measure, and (iii) practical KPIs that can be reported without a full-scale LCA study. In this context, “energy consumption” is intended as a step-resolved quantity (e.g., scanning vs. data processing vs. simulation vs. manufacturing/repair vs. verification) that can be reported per step and, where needed, aggregated to a workflow total, to avoid ambiguity between “single-stage” and “full-chain” accounting (see ISO 14955-1:2017 [108] and ISO 14955-2:2018 [109]).

Life Cycle Assessment (LCA) provides the standardized framework to define goal and scope, compile a Life Cycle Inventory (LCI), perform impact assessment (LCIA), and interpret results (see ISO 14040:2006 [110] and ISO 14044:2006 [111]). However, within RE workflows, a full LCA can be data-intensive and highly sensitive to system boundaries and allocation choices, especially when iteration/reworking loops vary across cases. Therefore, the proposed inventory and KPI set is intended as a lightweight and reproducible screening/complementary approach: it supports consistent, step-resolved reporting of dominant energy/resource/time drivers across the scan–model–simulate–manufacture chain and can serve as an LCI-ready backbone that may be expanded into a full LCA study when needed (see ILCD guidance [112]).

For repair routes involving laser/DED or related processes, reusable unit-process life-cycle inventory (UPLCI-style) models offer a structured way to report energy/resource drivers that are comparable across parts and settings [113]. Recent work on advanced laser-cladding variants (e.g., magnetic field-assisted laser cladding of high-entropy alloy coatings) further shows how process conditions can refine microstructures and reduce residual stresses, directly impacting coating integrity and durability—factors that should be reflected in chain-level sustainability and performance reporting [114]. This unit-process logic is consistent with ISO-based LCA/LCI practice and ILCD guidance for building consistent LCI [110,112]. At the process-chain level, recent reviews on laser-cladding-based repair/remanufacturing also motivate why sustainability should be discussed as a chain (scan → rebuild → machine → verify), not as a single process snapshot [115].

The four headline metrics (material loss, energy consumption, reworking, and lead time) can be quantified with a small set of shop-floor and digital-thread records when reported per chain step, and Table 9 operationalizes their minimum measurement needs and reporting format. For physical steps, energy should be captured from machine-integrated power logs or external power meters following standardized measurement principles for machine tools [108,109], then aggregated to a workflow total; where useful, monitoring models can further decompose electricity demand into baseline/constant and operation-dependent contributions [116], and empirical studies report step-wise electricity demand across additive manufacturing equipment [117,118]. For digital steps (data processing and simulation), workstation energy can be measured directly or estimated from hardware telemetry (e.g., CPU package energy/power) multiplied by wall-time [119,120]; when GPU readings are used (e.g., via NVIDIA-smi), known sampling/representativeness caveats should be acknowledged and mitigated [121]. Material loss can be reported as removed/scrapped mass (e.g., pre/post mass difference or collected chip/swarf mass) and, for additive repair, as deposited feedstock mass versus net added mass/yield. Reworking can be expressed as an iteration count (re-scan/rebuild/redo loops) and, when possible, as the associated additional time/energy/material beyond the first-pass route. Lead time should include both processing time and waiting/transport/queue times so that iteration-driven delays are visible; KPI definitions and reporting conventions can be aligned with manufacturing operations management standards (see ISO 22400-2:2014 [122]).

Table 9 summarizes the proposed sustainability framework for the scan–model–simulate–manufacture chain by linking each step to its main sustainability lever, the minimum inventory items to collect, and a small set of practical KPIs that can be reported consistently across case studies.

Table 9. Sustainability levers and a minimal inventory/KPI set across the scan–model–simulate–manufacture chain, structured as a two-scenario baseline (repair/remanufacture vs. replacement) to support consistent, step-resolved and decision-grade comparisons.

Chain Step	Sustainability Lever (Mechanism)	What to Measure (Inventory Items)	Practical KPIs (Examples)	Typical Data Sources	Notes/Pitfalls	Ref.
Scan (acquisition + planning)	Reduce re-scans and downstream scrap by task-driven scan planning and traceable settings (coverage, incidence, fixturing).	Number of scans/stations; scan time; energy of scanner/robot; auxiliary materials (e.g., spray coating mass); travel/logistics if on-site.	kWh per acquired part; # acquisition iterations; coating mass (g); % coverage on functional interfaces.	Machine power logs; robot cycle logs; scan reports; operator logs; travel distance.	Coatings can shift dimensions; report thickness assumptions. Energy is often small vs. repair, but iterations dominate.
Model (reconstruction + CAD editing)	Avoid over-processing (smoothing/repair/defeature) that triggers redesign loops; keep raw-to-processed traceability.	CPU/GPU time (optional); number of reconstruction iterations; parameter settings (filters, tolerances); manual editing time.	Wall-time per iteration; # reworking loops; tolerance settings (mm) tied to task.	Software logs; version control; engineering change logs.	Computation footprint is usually secondary, but the ‘iteration count’ is a strong proxy for waste and delay.	[123]
Simulate (mesh + BCs + uncertainty)	Use simulation to choose a repair scope that meets requirements the first time; quantify sensitivity to geometry/BCs to prevent over-conservatism.	Meshing time; solver runs; surrogate/model order reductions; uncertainty bands on QoIs (stress, life, stiffness).	# solver iterations; uncertainty band width (%); probability of meeting spec.	CAE logs; solver reports; DoE/sensitivity analyses; validation tests.	Poor BC transfer can dominate errors, causing reworking. Capture BC assumptions explicitly.	[10,124]
Manufacture/repair (process route)	Maximize lifetime extension per unit impact by selecting the right repair route and minimizing deposited/removed material.	Energy and gas consumption; feedstock mass; shielding gas; consumables; pre/post-machining; heat treatments; scrap rates.	kg CO2e per repaired part; material yield (%); deposited mass (g); kWh per repair.	Machine power meters; process monitoring; material certificates; LCI databases; supplier data.	Electricity mix and feedstock production often dominate impacts; include transport if material is sourced remotely.	[105,106,113]
Verify + release (inspection + documentation)	Prevent hidden defects and premature failures; verification avoids future scrap and downtime.	Inspection time; energy; consumables; reworking fraction; pass/fail statistics.	% first-pass yield; # reworking loops; inspection time (min).	Metrology logs; NDT reports; quality systems.	Verification needs to align with failure mode; avoid ‘inspection for its own sake’ by defining acceptance criteria early.	[115]
Cross-cutting: baseline and lifetime extension	Compare repair/remanufacturing against a clear reference scenario (replace with new) using a consistent functional unit.	Functional unit; system boundaries; lifetime extension; allocation rules; impact method; electricity mix.	Δ kg CO2e vs. replacement; CO2e per year of additional service; cost per year (optional).	LCA goal/scope; process chain inventory; supplier and grid data; cost accounting (if LCC).	Results are sensitive to assumed lifetime, utilization, and electricity mix—report them explicitly.	[107]

Table 9. Sustainability levers and a minimal inventory/KPI set across the scan–model–simulate–manufacture chain, structured as a two-scenario baseline (repair/remanufacture vs. replacement) to support consistent, step-resolved and decision-grade comparisons.

Chain Step	Sustainability Lever (Mechanism)	What to Measure (Inventory Items)	Practical KPIs (Examples)	Typical Data Sources	Notes/Pitfalls	Ref.
Scan (acquisition + planning)	Reduce re-scans and downstream scrap by task-driven scan planning and traceable settings (coverage, incidence, fixturing).	Number of scans/stations; scan time; energy of scanner/robot; auxiliary materials (e.g., spray coating mass); travel/logistics if on-site.	kWh per acquired part; # acquisition iterations; coating mass (g); % coverage on functional interfaces.	Machine power logs; robot cycle logs; scan reports; operator logs; travel distance.	Coatings can shift dimensions; report thickness assumptions. Energy is often small vs. repair, but iterations dominate.
Model (reconstruction + CAD editing)	Avoid over-processing (smoothing/repair/defeature) that triggers redesign loops; keep raw-to-processed traceability.	CPU/GPU time (optional); number of reconstruction iterations; parameter settings (filters, tolerances); manual editing time.	Wall-time per iteration; # reworking loops; tolerance settings (mm) tied to task.	Software logs; version control; engineering change logs.	Computation footprint is usually secondary, but the ‘iteration count’ is a strong proxy for waste and delay.	[123]
Simulate (mesh + BCs + uncertainty)	Use simulation to choose a repair scope that meets requirements the first time; quantify sensitivity to geometry/BCs to prevent over-conservatism.	Meshing time; solver runs; surrogate/model order reductions; uncertainty bands on QoIs (stress, life, stiffness).	# solver iterations; uncertainty band width (%); probability of meeting spec.	CAE logs; solver reports; DoE/sensitivity analyses; validation tests.	Poor BC transfer can dominate errors, causing reworking. Capture BC assumptions explicitly.	[10,124]
Manufacture/repair (process route)	Maximize lifetime extension per unit impact by selecting the right repair route and minimizing deposited/removed material.	Energy and gas consumption; feedstock mass; shielding gas; consumables; pre/post-machining; heat treatments; scrap rates.	kg CO2e per repaired part; material yield (%); deposited mass (g); kWh per repair.	Machine power meters; process monitoring; material certificates; LCI databases; supplier data.	Electricity mix and feedstock production often dominate impacts; include transport if material is sourced remotely.	[105,106,113]
Verify + release (inspection + documentation)	Prevent hidden defects and premature failures; verification avoids future scrap and downtime.	Inspection time; energy; consumables; reworking fraction; pass/fail statistics.	% first-pass yield; # reworking loops; inspection time (min).	Metrology logs; NDT reports; quality systems.	Verification needs to align with failure mode; avoid ‘inspection for its own sake’ by defining acceptance criteria early.	[115]
Cross-cutting: baseline and lifetime extension	Compare repair/remanufacturing against a clear reference scenario (replace with new) using a consistent functional unit.	Functional unit; system boundaries; lifetime extension; allocation rules; impact method; electricity mix.	Δ kg CO2e vs. replacement; CO2e per year of additional service; cost per year (optional).	LCA goal/scope; process chain inventory; supplier and grid data; cost accounting (if LCC).	Results are sensitive to assumed lifetime, utilization, and electricity mix—report them explicitly.	[107]

6.4. Practical Reporting Checklist for Decision-Grade Claims (Reproducibility + Sustainability)

Table 10. Practical reporting checklist for scan–model–simulate–manufacture studies, listing the minimum metadata to make sustainability claims reproducible and decision-grade by ensuring a clear baseline (functional unit and boundaries), traceability from raw scans to released parts, and explicit reporting of key process inventories and assumptions across the chain.

Reporting Block	Minimum Items to Report (Practical Checklist)	Why It Matters for Sustainability/Decision-Grade Claims	Examples/Notes
Goal and functional unit	Define the decision question (e.g., repair vs. replace); functional unit (e.g., 1 component meeting spec); reference scenario; assumed lifetime extension.	Without a baseline and functional unit, sustainability claims are not comparable and can be misleading.	Consider reporting ‘CO2e per additional service-year’ when lifetime extension is the key benefit [107].
System boundary and allocation	Specify included stages (scan, modeling, simulation, repair, post-processing, verification, transport); allocation rules for shared equipment; electricity mix.	Boundaries and electricity mix can flip conclusions for energy-intensive routes.	State the grid mix (country/region and year) and whether renewable sourcing is contractual [106].
Acquisition (scan)	Sensor type; calibration status; surface prep/coating type and assumed thickness; view planning rationale; number of views/stations; scan time.	Iteration count and surface prep can dominate time/cost and bias geometry; clear reporting supports reproducibility.	If a coating is used, report whether thickness compensation was applied and how.
Registration and reconstruction	Registration strategy (targets vs. targetless); acceptance criteria; filtering/denoising settings; reconstruction tolerance; versioning (raw vs. processed).	These choices control geometry fidelity and the number of reworking loops; they also determine what is ‘structural truth’ later.	A ‘raw-to-processed’ archive supports traceability when validation fails.
Simulation (scan-to-simulation)	Meshing strategy; element type/size; defeaturing rules with thresholds; BC transfer method; material model assumptions; QoIs and uncertainty/sensitivity method.	Avoids over-claiming: mechanical outputs should be tied to model assumptions and their uncertainty.	Report which features were removed and why (QoI-based).
Repair/manufacturing route	Process route (e.g., L-DED, WAAM repair, laser cladding + machining); feedstock type; deposited/removed mass; energy and gas; key parameters; yields/scrap.	This is typically the dominant contributor to impacts; reporting enables meaningful LCA/LCC and comparison across routes.	UPLCI-style inventories can help structure data [113].
Verification and acceptance	Inspection/NDT method; acceptance criteria; pass/fail and reworking rates; dimensional verification on functional interfaces; documentation for release.	Verification closes the loop: it links sustainability to reliability (avoiding premature scrap and downtime).	Report first-pass yield and major reworking causes if available.
Impact reporting (if quantified)	Impact method (e.g., GWP100); inventory sources/databases; uncertainty ranges; scenario analysis (electricity mix, transport, lifetime).	Small differences may not be meaningful without uncertainty; scenario analysis improves robustness.	For additive repair, material/feedstock often dominates: report powder/wire provenance [105].
Transparency and data reuse	Provide a minimal dataset summary (energy, material, time) and metadata sufficient for reuse; declare software versions.	Facilitates meta-analysis and makes the review ‘citable’ as a reporting reference.	Digital life-cycle management concepts support consistent data capture across the chain [123].

provides a practical reporting checklist that specifies the minimum metadata required to make scan-to-manufacture sustainability claims reproducible and decision-grade, including functional unit and baseline definition, system boundaries, key process inventories (energy/material/yield), and traceability from raw scans to released parts.

A frequent reason why sustainability discussions are judged to be “weak” in manufacturing reviews is not a lack of ambition, but a lack of minimum metadata: unclear functional unit, missing boundaries, unreported yields, unknown electricity mix, and no traceability from raw scans to final decisions. Table 10 provides a compact checklist that can be applied whether the paper reports a full LCA, a partial inventory, or only qualitative sustainability implications.

The checklist is consistent with how digital life-cycle management and digital-thread concepts are used to support consistent data capture across manufacturing activities [123]. It also aligns with the broader “life-cycle engineering embedded in design/manufacturing choices” perspective increasingly emphasized in recent design-for-AM and sustainability syntheses [124,125].

From an interpretation standpoint, the step-wise inventory enables hotspot attribution at the chain level: in some cases, the dominant driver is iteration/reworking (calling for more robust acquisition, registration, and verification gates), while in others it is material yield/feedstock or the electricity mix (calling for route and process optimization). When conclusions appear boundary-sensitive (functional unit, lifetime extension, allocation), the same records provide an LCI-ready backbone that can be expanded into a full LCA with transparent assumptions.

7. Open Challenges and Research Directions for Robust, Automatable, and Sustainable RE in Metal Manufacturing

Despite rapid progress in sensing, geometry processing, and simulation toolchains, industrial reverse engineering (RE) for metal manufacturing still faces recurring bottlenecks that prevent fully reliable automation. The core implication is that “good geometry” is not sufficient: industrial adoption requires decision-grade outputs whose uncertainty, provenance, and downstream impact on tolerances and QoIs can be audited and defended. The most persistent limitations are system-level rather than single-step: weak traceability from measurement to model, limited interoperability across the digital thread, and insufficient evidence that scan-derived models remain decision-grade when used for mechanical decisions (fatigue, contact, repair planning, acceptance/rejection). Accordingly, the open challenges below are framed as adoption barriers, each linked to measurable evaluation outputs and qualification-oriented evidence. The main open challenges are mapped to actionable research directions and evaluation metrics in

Table 11. Open challenges in industrial RE for metal manufacturing mapped to actionable research directions and decision-relevant evaluation metrics to connect each bottleneck to measurable validation targets suitable for qualification and automation.

Challenge/Bottleneck	Why It Matters (Metal Manufacturing)	Research Direction	Evaluation Metrics/Reporting	Ref.
Metrology-grade benchmarks and traceable ground truth for industrial RE	Without traceable reference geometry and GD and T ground truth, it is hard to compare pipelines, quantify bias (e.g., edges/filets), or certify scan-to-model results.	Create benchmark parts + datasets with calibrated reference (CMM/CT), multi-sensor scans, and feature-level annotations (GD and T, radii, wall thickness). Publish protocols for repeatability and uncertainty budgets.	Feature-level deviation (GD and T), edge/filet radius error, thickness error, repeatability (σ), uncertainty budget, pass/fail consistency vs. tolerance.	[60,126]
Robust, quantified registration in cluttered shop-floor conditions	Small misalignments shift hole axes and filet positions, biasing interface fits and fatigue hotspots; errors are non-uniform and task-dependent.	Registration with explicit uncertainty models, outlier-robust alignment, and feature-/datum-aware objectives aligned with manufacturing intent (datums, functional interfaces).	Residual distributions (not only RMSE), datum/feature drift, uncertainty bounds, robustness to occlusion/surface finish changes, and runtime per scan.	[55]
Automation that preserves sharp features under noise and incomplete coverage	Edges, chamfers, and small radii control stress concentration and assembly fit; overly smooth reconstructions hide critical detail.	Feature-preserving reconstruction/repair with controllable priors; hybrid geometric + learning methods that explicitly preserve feature curves and primitives.	Edge/curve recall, radius preservation error, Hausdorff distance near features, topology validity, and downstream meshing success rate.	[65,78]
From scans/meshes to editable CAD (B-rep/parametric) suitable for manufacturing changes	Repair and redesign require constraint-aware CAD edits (holes, bosses, blends) rather than static meshes; current pipelines often stop at “watertight mesh”.	Semantic reconstruction that outputs B-rep/feature graphs, supports constraints, and provides confidence measures for recovered features.	Editability tests (feature constraints), B-rep validity, feature recognition accuracy, tolerance consistency after edits, and CAD regeneration success.	[77,127,128]
Defect detection/segmentation that generalizes across metals, finishes, and processes	Automated decisions (accept/repair/remanufacture) require reliable defect localization on reflective/rough surfaces and complex weld/AM geometries.	Domain-general defect models with physics-informed priors; combine point clouds with process/thermal history and uncertainty-aware thresholds.	Detection F1/IoU, false-alarm cost, sensitivity to surface condition, calibration of confidence, and cross-site generalization tests.	[129,130]
Boundary-condition and load-path transfer from scan/CAD to simulation models	Simulation-ready models require correct contacts, constraints, preload, and datum definitions; mis-specified BCs can dominate mechanical error.	BC inference using feature semantics + assembly context; standardized mapping from MBD/PMI to CAE; and automated contact detection with verification checks.	BC consistency checks, contact completeness, QoI sensitivity to BC perturbations, and traceable mapping from datums/PMI to CAE entities.	[126,131]
Propagation of geometric uncertainty to quantities of interest (QoIs)	Geometry uncertainty is spatially heterogeneous; QoIs (fatigue hotspot stress, contact pressure) are highly non-linear and locally sensitive.	Uncertainty-aware scan-to-simulation (UQ) with local sensitivity maps; link measurement uncertainty to probabilistic response envelopes.	QoI error bounds vs. experiments, local sensitivity maps, uncertainty propagation method, validation cases, and computational cost.	[126,132]
Multi-scale modeling for repaired components (macro geometry defects ↔ microstructure-sensitive/anisotropic repair zones)	Service-life predictions for repaired parts depend on both scan-captured macroscopic deviations (distortion/warpage, local shape errors) and heterogeneous, often anisotropic material behavior in AM/DED repair deposits. Assuming homogeneous/isotropic properties can mispredict fatigue hotspots, crack initiation sites, and durability, undermining acceptance decisions for high-value components.	Adopt process–structure–property/ICME-type workflows that: (i) transfer scan-derived geometry and distortion/residual-stress fields into the structural model; (ii) infer repair-zone quality descriptors from process history and/or characterization (porosity/LoF indicators, texture proxies); (iii) assign spatially varying (possibly anisotropic) constitutive and fatigue models; and (iv) propagate uncertainty across scales with validation against benchmark-quality datasets.	Report the defect-to-property mapping assumptions, spatial resolution of property fields, and calibration datasets; quantify sensitivity of fatigue-relevant QoIs to anisotropy/defect parameters; compare predicted vs. measured distortion/residual stress and (where available) microstructure descriptors; report QoI errors (e.g., life, crack-initiation location) and uncertainty-bound coverage.	[133,134,135,136,137,138,139]
Interoperability in the digital thread (scan ↔ CAD/PMI ↔ CAE ↔ manufacturing)	RE outputs often break when moving across tools; loss of datums/PMI and versioning undermines traceability and automation.	Adopt and extend exchange frameworks (e.g., QIF/MBD) for point clouds, feature semantics, and uncertainty; develop test suites for interoperability.	Information loss audits (datums/PMI), round-trip integrity tests, provenance tracking, and audit logs for revisions and decisions.	[60,131]
Qualification and certification of AI-assisted RE pipelines	High-value components (aerospace/energy) need auditable evidence that AI decisions are reliable under shift; black-box outputs are hard to certify.	Hybrid verification (rules + learning), calibration of model confidence, and standardized test protocols for domain shift and failure modes.	Calibration error, out-of-distribution detection, traceable failure cases, robustness under shift, and human-in-the-loop intervention rate.	[1,129]
RE-driven repair planning and closed-loop remanufacturing	Repair chains need consistent geometry capture, damage volume extraction, toolpath planning, and verification—often across multiple iterations.	Closed-loop scan–plan–deposit-inspect with simulation-assisted planning; quantify how scan uncertainty propagates to repair bead/thermal distortion and final geometry.	Repair volume extraction error, toolpath feasibility, as-repaired deviation, re-scan iteration count, and process time and scrap avoided.	[1,140,141]

, while

Table 12. Compact benchmarking and reporting protocol for automatable scan-to-model-to-simulation workflows, defining representative tasks, required inputs/outputs, minimum metadata, and primary metrics to enable comparable and auditable evaluations across pipelines.

Benchmark Task	Inputs → Outputs	What to Publish (Minimum)	Primary Metrics	Ref.
Acquisition planning and coverage	Part + constraints → viewpoint/trajectory plan + coverage map	CAD of benchmark part; scan poses/trajectory; environmental conditions; coating thickness policy; raw + processed scans	Coverage %, incidence-angle distribution, time/cycle time, repeatability (σ) across runs	[60,126]
Registration (target-based and target-less)	Partial scans → aligned point cloud/mesh + uncertainty map	Datum/target definitions; overlap %; residual distributions; acceptance thresholds; raw vs. filtered alignment results	Feature drift, residual histograms, robustness to occlusion/outliers, runtime	[55]
Feature extraction and manufacturing feature recognition	Point cloud/mesh/B-rep → primitives/features + confidence	Feature labels (holes, pockets, blends); tolerance classes; confusion matrices; failure cases	Precision/recall per feature; topology validity; editability success	[142,143]
B-rep/editable CAD reconstruction	Point cloud/mesh → B-rep/parametric CAD + constraints	Neutral CAD (STEP) + constraints; reconstruction logs; regeneration tests in CAD; validity checks	B-rep validity, constraint satisfaction, deviation at functional interfaces, CAD regeneration rate	[78,127,128]
Mesh repair and feature-preserving simplification	Noisy mesh → watertight mesh + feature curves	Healing thresholds; hole-filling policies; before/after mesh statistics; feature-curve preservation measures	Watertightness, topology defects removed, edge preservation error, meshing success	[65,78]
Defect detection and segmentation for repair decisions	Scan/mesh + context → defect regions + severity	Ground-truth defect labels; surface condition; false-alarm cost model; cross-site tests	IoU/F1, false-positive cost, calibration, generalization under surface/process shift	[129,130]
Scan-to-simulation uncertainty propagation	Geometry + uncertainty → QoI distribution (stress, fatigue, contact)	Uncertainty model; QoIs and load cases; experimental validation; sensitivity maps	QoI error bounds, sensitivity localization, computational cost, and reproducibility	[126,132]
Multi-scale repair zone modeling and service-life prediction (macro defects + microstructure/anisotropy)	Inputs: scan-derived geometry (and deviation map), identified critical regions, repair process history (toolpath/parameters), and/or targeted microstructure data (e.g., CT porosity, EBSD texture). Output: structural model with spatially varying (potentially anisotropic) repair-zone properties and predicted fatigue-relevant QoIs (life, hotspot location, crack-growth metrics) with uncertainty bounds.	Aligned scan + reference geometry; deviation/distortion map and (if available) residual-stress/deflection ground truth; repair process history (or equivalent metadata); repair-zone characterization (porosity/LoF proxies, texture/microstructure descriptors) and coupon-level mechanical data; simulation inputs (material field definition) + scripts for reproducibility.	QoI error vs. experiments (life/hotspot/crack initiation); sensitivity of QoIs to anisotropy/defect parameters; agreement on distortion/residual stress where available; uncertainty-bound coverage; computational cost (time/mesh size) at stated accuracy.	[134,136,137,138,139]
Digital thread exchange (scan–CAD/PMI–CAE–manufacturing)	Model versions → exchanged artifacts with provenance	Exchange format versions; datums/PMI retention checks; audit trail of revisions	Round-trip integrity, info-loss score, traceability completeness	[60,131]
Closed-loop repair chain (scan-plan-deposit-inspect)	Damaged part → repaired part + verification report	Damage definition; repair volume extraction; toolpaths; in-process data; post-repair metrology and acceptance criteria	As-repaired deviation, iteration count, process time, reworking rate	[140,141]

proposes a compact benchmarking and reporting protocol designed specifically for automatable scan-to-model-to-simulation workflows.

7.1. Metrology-Grade Benchmarks and Traceable Ground Truth (Still the Main Bottleneck)

A core limitation for robust automation is the lack of metrology-grade benchmark datasets that reflect industrial conditions (reflective metals, mixed finishes, partial occlusions, fixture variability) while providing traceable ground truth at the feature level (GD and T-relevant references, radii, wall thickness, datums). Recent literature in industrial point-cloud AI repeatedly highlights that many learning-based pipelines are trained/evaluated on datasets that do not fully capture shop-floor variability, which complicates reproducibility and qualification for high-value components [60]. From a metrology perspective, uncertainty is task-specific and cannot be reduced to a single “scanner accuracy” number; recent work explicitly frames uncertainty assessment for dense point clouds as an enabling condition for reliable industrial usage [126].

Industrial implication: Without traceable feature-level ground truth, pipelines cannot be qualified against acceptance decisions (pass/fail vs. tolerance) and therefore remain “research-grade” even when geometric RMSE appears low.

Research direction: Benchmark parts and datasets should be designed around manufacturing intent, not geometry alone: (i) traceable reference geometry, (ii) multi-sensor acquisitions, (iii) feature-level labels/tolerance classes, and (iv) explicit documentation of coatings, environment, and fixturing. The qualification-relevant outcome is decision stability: repeatable pass/fail outcomes and bounded QoI variation under realistic acquisition variability, not only in RMSE terms, but in terms of pass/fail stability vs. tolerances and mechanical decision outcomes (see Table 11 for the corresponding evaluation metrics).

7.2. Registration That Is Robust and Quantified (Not Only “Best-Fit”)

Registration remains a dominant failure mode in industrial RE because misalignment error is rarely uniform: it concentrates where the geometry is weakly constrained, where occlusions are frequent, or where surface quality degrades point reliability. This motivates growing interest in surveys that classify registration families and clarify evaluation practices, including recent consolidations that emphasize the link between algorithm assumptions and robustness [55].

Industrial implication: “best-fit” alignment can be mechanically misleading—small datum/axis shifts may be negligible in global deviation maps but critical for GD and T features and BC transfer. For manufacturing-driven workflows, the open challenge is therefore to output alignment together with reliability/uncertainty information that identifies where alignment is trustworthy enough for tolerancing or for transferring loads/contacts into CAE.

Research direction: registration objectives should increasingly become datum- and feature-aware, aligning “best-fit” logic with functional interfaces rather than minimizing global residuals. Success criteria should include feature drift bounds (axes, planes, interfaces) and their effect on downstream QoIs, as reflected in Table 11 and Table 12.

7.3. Feature Preservation Under Noise, Incomplete Coverage, and Repair Operations

Industrial metal components often contain small filets, chamfers, and transitions that are geometrically subtle but mechanically dominant. A persistent open problem is that automated reconstruction and repair operations may over-smooth these details or alter them unpredictably, especially when the scan contains holes, outliers, or mixed-quality regions. Recent work on CAD-focused reconstruction from point clouds explicitly targets feature preservation and CAD-like output consistency (e.g., eCAD-Net [78]), while contemporary mesh repair research is increasingly explicit about preserving sharp features during “make-it-watertight” steps [65].

Industrial implication: Uncontrolled feature erosion translates into unquantified shifts in stress raisers and contact conditions, undermining fatigue and sealing decisions even when global deviation remains acceptable.

Research direction: Hybrid pipelines that combine (i) geometric priors (primitives/feature curves) with (ii) learning-based inference should be evaluated against feature fidelity and mechanics-relevant consequences, not only on surface distance metrics, but on filet/radius fidelity, edge location stability, and hotspot migration in stress/contact fields. These priorities are captured in Table 11 and translated into benchmarking tasks/metrics in Table 12.

7.4. From Scan-to-CAD to Editable CAD: Semantics, Constraints, and Manufacturing Intent

For metal manufacturing, a mesh that visually matches the part is often insufficient. Repair, remanufacturing, and redesign require editable CAD with semantic features (holes, blends, pockets) and constraint consistency. Recent work in CAD reconstruction and machining-feature reasoning shows strong momentum toward representation learning on B-reps and feature graphs (e.g., Brep2Seq [127]; BRepGAT [128]), while broader surveys synthesize deep-learning approaches for 3D CAD reconstruction and clarify what remains unsolved in terms of editability and validity guarantees [77]. From an application angle, learning-based manufacturing feature recognition from CAD/B-rep structures is increasingly mature and points to practical integration pathways for CAM/repair planning (e.g., DeepFeature [142]; AAGNet [143]).

Industrial implication: Without constraint-aware editability and validity guarantees, downstream CAM/CAE modifications become brittle and non-auditable, preventing safe automation in repair chains.

Research direction: Industrial adoption requires confidence-aware semantic reconstruction, i.e., reporting which reconstructed features are reliable, which are inferred, and which require human confirmation, so that CAD edits and downstream simulation do not silently rely on unstable geometry. Table 11 summarizes the challenge as “editable CAD suitable for manufacturing changes”, and Table 12 proposes qualification-oriented editability tests (constraint regeneration and CAD validity checks) as standard evaluation outputs.

7.5. Defect Detection and Segmentation That Generalizes Across Metals, Finishes, and Processes

Automatable RE in metal manufacturing increasingly intersects with defect detection (wear, cracks, lack-of-fusion indications on repaired surfaces, geometric deviations in weldments). A major difficulty is domain shift: defects and “false defects” appear different across alloys, surface finishes, lighting, and scanning principles. Recent synthesis work on industrial 3D defect classification/segmentation emphasizes both progress and persistent gaps in generalization and benchmark consistency [129]. At a more applied level, recent work using point clouds to detect geometric defects (e.g., gear manufacturing) illustrates the feasibility of production-focused pipelines but also highlights the need for robust labeling and cross-condition validation [130].

Industrial implication: Generalization failures directly translate into economic risk (false alarms → unnecessary reworking/scrap; missed defects → unsafe return-to-service), so defect detection must be evaluated with cost/risk-aware metrics rather than accuracy alone.

Research direction: Future pipelines should integrate defect detection with uncertainty-aware thresholds and decision-cost models (false-alarm cost vs. missed-defect risk). This is treated explicitly in Table 11 and translated into a benchmark task template in Table 12 (“Defect detection and segmentation for repair decisions”).

7.6. Scan-to-Simulation Trust: Boundary Condition Transfer and Uncertainty Propagation to QoIs

In many industrial cases, the dominant simulation error is not the mesh density but boundary condition (BC) specification: contacts, constraints, preload, and datum choices. When models originate from scans, the BC definition becomes even more fragile unless the RE pipeline preserves semantic features and interface information. A parallel challenge is that geometric uncertainty is spatially heterogeneous, while mechanical quantities of interest (QoIs) are strongly localized (fatigue hotspots, contact pressures). Recent work in precision engineering proposes scanner digital twins and structured methods for error mapping, which can support more defensible uncertainty statements at the acquisition level [132], while metrology-focused studies further motivate explicit uncertainty reporting as a prerequisite for trustworthy downstream decisions [126].

Industrial implication: Without BC traceability and uncertainty-to-QoI propagation, scan-to-simulation outputs cannot be used as acceptance evidence, because the dominant variability is “hidden” and not bounded.

Research direction: Scan-to-simulation should increasingly be evaluated as an uncertainty-to-QoI pipeline, where the goal is not only a nominal stress field, but a bounded QoI envelope. Table 11 summarizes the challenge and proposes QoI-bound validation metrics, and Table 12 provides a compact reporting structure for uncertainty propagation studies.

7.7. Multi-Scale Modeling: Coupling Macroscopic Geometry Defects with Microstructure-Sensitive Properties in Repair Zones

A further open challenge concerns multi-scale modeling, namely the consistent coupling of macroscopic as-built geometric deviations captured by scanning (e.g., distortions, warpage, local shape errors) with microstructure-dependent material behavior in repaired/additively deposited zones. Multi-scale approaches for additive manufacturing have been widely discussed, highlighting the need to bridge part-scale thermo-mechanical fields with microstructure evolution and local properties [133]. In this direction, integrated process–structure–property (ICME-type) frameworks have been proposed to link process thermal history to microstructure and, ultimately, to mechanical response [134]. However, even when part-scale models predict residual stress/distortion efficiently, transferring this information into reliable, location-specific constitutive properties remains non-trivial [135]. This limitation is particularly relevant in DED/AM repair deposits, where heterogeneous microstructures and anisotropy driven by thermal history, texture, and defect populations can strongly affect mechanical response and fatigue performance; experimental evidence of anisotropic properties in DED-processed alloys supports the need for spatially varying and direction-dependent material descriptions in repaired zones [136]. Process variants in laser cladding (e.g., magnetic field assistance) have been shown to reduce residual stress and modify microstructure, reinforcing the need to couple scan-derived geometry with microstructure-sensitive properties for life prediction [114]. Repair-focused evidence likewise shows that processing conditions strongly affect microstructure, porosity, and mechanical properties, so homogeneous/isotropic assumptions can be misleading for life prediction [137].

Industrial implication: Assuming homogeneous/isotropic properties in repaired zones can produce non-conservative life predictions, so qualification requires explicit coupling (or bounding) of geometry defects with property variability in the repaired volume.

Research direction: Robust scan-to-simulation workflows should evolve toward defect–microstructure coupling strategies in which scan-derived macroscopic geometry/defects are transferred to structural models, and repaired zones receive spatially varying (possibly anisotropic) constitutive/fatigue properties inferred from process history and/or targeted characterization prior to fatigue/fracture assessment. This coupling is critical because defect- and microstructure-driven fatigue life prediction approaches explicitly show the sensitivity of life estimates to defect populations and local material conditions [138]. Progress is currently constrained by calibration effort, uncertainty propagation across scales, and the scarcity of benchmark datasets that jointly provide geometry, distortion/residual stress, and microstructure/property ground truth; AM benchmark initiatives provide useful exemplars of openly accessible datasets that can support validation and uncertainty quantification [139].

7.8. Interoperability and the Digital Thread: Preventing Information Loss from Scan to Manufacturing Action

Even when individual tools perform well, industrial workflows often fail at interfaces: point clouds lose provenance, datums disappear during format conversion, PMI/MBD information is not preserved, and simulation assumptions are not traceable back to the measurement chain. Recent work on Quality Information Framework (QIF) in the digital thread explicitly addresses model-based quality integration and interoperability gaps that are directly relevant to RE-enabled manufacturing decisions [131]. In parallel, industrial point-cloud AI surveys stress that “deployment readiness” depends heavily on data management, provenance, and consistent representation choices—issues that are not solved by model accuracy alone [60].

Industrial implication: Interoperability failures convert technical uncertainty into organizational risk—decisions cannot be audited end-to-end, which blocks certification and cross-site repair networks.

Research direction: Beyond adopting exchange standards, the field needs round-trip integrity tests that quantify what is lost (and where) and how that loss affects decisions. These integrity checks are captured as a benchmark task and reporting template in Table 12.

7.9. Qualification of AI-Assisted RE Pipelines and Closed-Loop Remanufacturing

For high-value metal components, automation must be auditable. Systematic reviews of RE for remanufacturing highlight persistent challenges in integration, tolerance handling, scanning limitations, and automation barriers, and explicitly point to editable B-rep reconstruction and AI-driven feature recognition as major future directions [1]. In repair chains, recent applied work illustrates the growing integration of 3D scanning with process planning and simulation—for example, in toolpath planning for deposition-based repair [140] and in scan + simulation workflows used to analyze geometric accuracy in welded components [141]. These contributions support a broader research push toward closed-loop “scan–plan–deposit–inspect” workflows, but they also emphasize that certification-grade automation requires robust validation, confidence calibration, and failure-mode documentation.

Industrial implication: “Automation” is acceptable only when failure modes are known, confidence is calibrated, and a traceable audit trail exists from measurement to released part—otherwise, the digital thread cannot support return-to-service decisions.

Research direction: Qualification should combine calibrated confidence reporting, domain-shift testing, and traceable audit trails linking measurement → reconstruction → simulation → manufacturing action. These requirements are summarized in Table 11 and operationalized in Table 12.

8. Conclusions

This review has framed reverse engineering (RE) for metal manufacturing as a scan–model–simulate–manufacture decision chain, where value is created only when geometry becomes decision-grade evidence—traceable, uncertainty-aware, and fit for downstream CAD/CAE/CAM and repair/remanufacturing actions. Section 2, Section 3 and Section 4 highlighted that acquisition, reconstruction, and scan-to-simulation are tightly coupled through uncertainty: surface–sensor interaction on metals, multi-view registration drift, and XCT surface determination choices can propagate into feature-level bias; in turn, reconstruction tolerances, mesh healing/decimation, defeaturing/idealization thresholds, and boundary-condition transfer can selectively suppress or shift stress-driving details, especially at small radii, transitions, and contact interfaces.

A manufacturing-oriented contribution of this work is the consolidation of practical “knobs” and their structural consequences (Table 3 and Table 4), together with a repair/remanufacturing view where RE outputs enable explicit decision gates and verification loops (Table 7 and Table 8). This positions RE not as an isolated digitization task, but as a closed-loop workflow that supports repair-vs-replace screening, repair-volume definition, process-route selection, and qualification with re-inspection.

To keep sustainability actionable rather than aspirational, Section 6 proposed a concise framework that treats sustainability as a two-scenario comparison (RE-enabled remanufacture vs. replacement) and prioritizes the few variables that usually dominate conclusions—iterations, yields, energy mix, feedstock, and lifetime extension (Table 9), supported by a minimal reporting checklist designed for reproducibility and meta-analysis (Table 10). Section 7 then distilled the remaining barriers to robust automation into a small set of system-level gaps—metrology-grade benchmarks, uncertainty-to-QoI propagation, semantic/constraint-aware CAD reconstruction, and interoperable digital-thread artifacts—mapped to practical evaluation metrics and reporting templates (Table 11 and Table 12).

Overall, the field is approaching the point where automation is technically feasible, but widespread industrial adoption will depend on whether RE pipelines can provide auditable traceability, task-specific uncertainty, and standardized reporting/benchmarking that link geometric decisions to mechanical and manufacturing outcomes. At the same time, the review’s contributions should be read within clear boundaries: as a structured narrative review, it does not introduce new experimental benchmarks or a unified quantitative dataset, and the numerical values cited are literature-based order-of-magnitude anchors rather than universal performance guarantees. The consolidated “workflow knobs” and decision gates summarize recurring practice across the literature, but they do not replace component-specific qualification, nor do they remove the need for case-by-case validation of boundary conditions, material assumptions, and uncertainty-to-QoI propagation when safety- or certification-critical decisions are involved. In addition, the sustainability framework is intended as a step-resolved screening inventory; decision-grade environmental claims still require an explicit LCA/LCI goal-and-scope definition, boundaries/allocation rules, and scenario uncertainty. Accordingly, future research should prioritize feature-level, decision-relevant evaluation, so that scan-to-model performance is assessed not only by geometric error, but by its impact on tolerances, structural responses, and the sustainability of manufacturing decisions.