Hospital Ventilation Optimization: Balancing Thermal Comfort and Energy Efficiency in Nonlinear Building Dynamics

Abstract

Despite growing interest in AI-driven Heating, Ventilation, and Air Conditioning (HVAC) systems, existing approaches often rely on static control strategies or offline simulations that fail to adapt to real-time environmental changes, especially in high-risk healthcare settings. There remains a critical gap in integrating dynamic, physics-informed control with human-centric design to simultaneously address infection control, energy efficiency, and occupant comfort in hospital environments. This study presents an AI-driven ventilation system integrating BIM, adaptive control, and computational fluid dynamics (CFD) to optimize hospital environments dynamically. The framework features (1) HVAC control using real-time sensor datasets; (2) CFD-validated architectural interventions (1.8 m partitions and the pressure range at a return vent); and (3) patient flow prediction for spatial efficiency. The system reduces airborne pathogen exposure by 61.96% (159 s vs. 418 s residence time) and achieves 51.85% energy savings (0.19 m/s airflow) while maintaining thermal comfort. Key innovations include adaptive energy management, pandemic-resilient design, and human-centric spatial planning. This work establishes a scalable model for sustainable hospitals that manages infection risk, energy use, and occupant comfort. Future directions include waste heat recovery and lifecycle analysis to further enhance dynamic system performance.

1. Introduction

Hospital ventilation systems are essential in creating healthcare environments and should prioritize patient safety, staff well-being, and operational sustainability [1,2]. These systems must balance competing demands: maintaining high indoor air quality, ensuring interior climate optimization, minimizing energy consumption, and complying with stringent health standards [3,4]. Effective ventilation has been shown to significantly reduce the transmission of airborne pathogens, including those responsible for tuberculosis and COVID-19 [5,6], and proper pressure differentials in isolation rooms and operating theaters strengthen pathogen mitigation [7,8]. Strategic placement of air supply diffusers and exhaust grilles, for instance, can minimize particle dispersion from visitors to vulnerable patients [9], showing how system design directly impacts clinical outcomes. Hospitals present uniquely complex challenges for ventilation design compared to other building types. They house vulnerable populations, including immunocompromised patients, and require strict infection control protocols such as airborne isolation and pressure zoning. Additionally, hospitals operate continuously and experience highly variable occupancy patterns due to emergency admissions, shift changes, and procedural demands. These factors make hospitals an ideal and high-stakes environment for evaluating adaptive Heating, Ventilation, and Air Conditioning (HVAC) systems that must balance infection control, energy efficiency, and occupant comfort in real time.

These operational demands also contribute to the high energy intensity of hospitals, which must reconcile efficiency with regulatory compliance. Specifically, standards such as ASHRAE (American Society of Heating, Refrigerating and Air-Conditioning Engineers) and the WHO (World Health Organization) Air Quality Guidelines prescribe specific thresholds for air changes per hour, filtration levels, and temperature ranges, yet meeting these requirements often conflicts with sustainability goals [10]. Current ventilation design practices must therefore solve multidimensional optimization challenges, including ventilation dynamics distribution [5], indoor climate adjustment [11], and the integration of advanced control strategies [12]. Suboptimal ventilation system performance can directly endanger patient safety, increase public health risks, and impose significant economic burdens through extended hospital stays and outbreak containment costs [13,14]. These consequences highlight the urgent requirement for advanced ventilation solutions.

Three shortcomings hinder ventilation performance improvements. First, although reinforcement learning (RL) has demonstrated significant potential for HVAC control in recent studies (e.g., [15,16]), most implementations remain constrained by static building models or pre-simulated computational fluid dynamics (CFD) data, limiting real-time adaptability to dynamic occupancy patterns and space utilization changes [17,18]. Second, CFD simulations typically refine ventilation designs offline [19], without mechanisms for dynamic adjustment during facility operation. Third, although existing HVAC architectures often prioritize singular objectives, such as energy efficiency [20] or thermal comfort [19], they typically neglect the trade-offs required for effective pathogen mitigation, which is a limitation particularly evident in clinical settings [14]. This oversight becomes acute in high-risk environments where close caregiver-patient interactions generate dynamic airborne dispersion patterns that challenge conventional simulation approaches [14,21,22]. Recent studies confirm that traditional systems fail to address (1) the rapid spatial-temporal variability of bioaerosols [16], and (2) the competing demands of energy conservation versus infection control [15]. Hence, the central research problem in this study is how to design a hospital ventilation system that dynamically integrates real-time environmental data, adaptive control, and spatial optimization to simultaneously achieve infection control, energy efficiency, and occupant comfort. Such a system can overcome the limitations of static models, offline simulations, and single-objective HVAC strategies.

This study introduces a self-correcting BIM-integrated AI-CFD system architecture to overcome these limitations. The system integrates dynamic sensor data with Long Short-Term Memory (LSTM) networks to predict space utilization patterns, employs adaptive control algorithms for HVAC, and validates performance through simulations. By jointly adjusting ventilation dynamics, energy use, and airborne particulate exposure, the system architecture advances beyond static designs to enable responsive ventilation tailored to clinical demands. The research specifically investigates previously understudied transmission risks arising from caregiver and patient postures, providing new insights into particle dispersion dynamics in active treatment environments.

2. Background

Emerging techniques integrate data-driven models (e.g., machine learning (ML)) to reduce computational costs and maintain prediction accuracy [23]. This hybrid approach enables dynamic controls in smart buildings [24,25,26]. Specifically, regression models and neural networks for aerodynamic improvements provide a thorough understanding of established methods to inspire innovative designs [27,28]. In another example, by predicting ventilation distribution using trained ML models based on limited simulation data, people can reduce airborne pathogens and improve air quality [29]. ML algorithms like Genetic Algorithms or Particle Swarm Optimization can get the most out of design parameters (e.g., geometry and boundary conditions) [30].

Long Short-Term Memory (LSTM) networks are a powerful tool for modeling nonlinear dynamics and fully leveraging complex systems, particularly in HVAC and system control. Their ability to capture temporal dependencies and non-monotonic relationships makes them ideal for predicting airflow patterns and energy consumption trends [3,4,31]. In HVAC optimization, LSTM networks have demonstrated exceptional capability in thermal load forecasting when trained on building automation system data [31]. The learning architectures were integrated with model predictive control (MPC) frameworks, and the nonlinear system dynamics were based on historical sensor logs to enable adaptive setpoint optimization [31,32]. Integrated LSTM and reinforcement learning (e.g., Proximal Policy Optimization (PPO)) can handle multi-objective trade-offs (e.g., energy vs. air quality) in real time [20]. Nevertheless, challenges persist in generalizing LSTM models across diverse building layouts and scaling training for high-dimensional CFD-coupled systems [16,17,18].

Proximal Policy Optimization (PPO) is an RL algorithm and has been applied in fluid dynamics and built environment control due to its stability in handling high-dimensional, nonlinear development problems. The advantage lies in its ability to process real-time CFD-derived airflow data while maintaining policy update constraints that prevent destabilizing control shifts [16,17,18]. Recent studies demonstrate its effectiveness in multi-objective scenarios: Nguyen et al. [15] achieved 2–12% energy savings in hospital ventilation by coupling PPO with transient CFD simulations, in the process of maintaining thermal comfort bounds. The algorithm outperforms traditional Proportional-Integral-Derivative (PID) controllers in suppressing airflow shortcuts [33]. However, current implementations face challenges in complex geometries and turbulent flow regimes due to their multi-scale nature and stochasticity, leading to high variance in training data that undermines RL algorithm performance and robustness [22]. Using deep RL (including PPO) for dynamic thermal management in data centers reported a cooling effect with 8 K lower maximum temperature (307 K vs. 315 K) compared to manually optimized control policies [34]. This positions PPO as a transformative approach for adaptive fluid dynamics control, though its integration with high-fidelity simulations remains an active research frontier.

Data-driven CFD modeling together with ML can enhance learning corrections and offer insights into designing efficient HVAC systems for next-generation buildings [35]. A typical ML-based CFD simulation process includes pre-processing to define the problem domain, creating a mesh (discretization of the domain into small cells), and specifying boundary and initial conditions [36]. The steps apply numerical methods to solve discretized equations and use computational resources to iterate towards a stable solution. In the post-processing step, results are visualized (e.g., velocity fields, pressure distribution, temperature gradients), and data are analyzed to derive insights. The advantages include reducing the need for expensive physical prototypes, exploring fluid behavior that may be difficult to measure experimentally, and testing extreme conditions without physical risk.

Table 1. Features of AI-Driven HVAC Design and Systems.

Ref	Keywords	Relevancy	Features and Advantages	Limitations
[37]	BIM, Airflow Simulation, Machine Learning	High (Dynamic HVAC control)	-Real-time BIM-CFD integration Focus: Healthcare	High computational load compromises adaptability
[38]	Nonlinear Optimization, BIM, Pattern Identification	High (Energy efficiency)	-ML-driven adaptive airflow control Method: Genetic Algorithms	Requires extensive sensor data; Not for hospitals.
[39]	Pattern Identification, ML, Airflow Simulation	High (Pathogen mitigation)	-BIM + IoT for occupancy patterns Dataset: An office room	Complex calibration; Not for hospitals.
[40]	BIM, Airflow Simulation, ML	Medium (Urban scale)	-Urban-scale BIM-CFD coupling Application: Smart cities	Limited to pre-trained scenarios; Not for hospitals
[41]	Nonlinear Optimization, BIM, ML	Medium (District energy)	-ML for wind flow prediction Focus: Renewable integration	Needs high-resolution CFD data; For apartments, not for hospitals
[42]	Pattern Identification, ML, Airflow Simulation	High (Post pandemic design)	-RL for energy grids Tool: Python (3.13.7)-based ML	Generalizes micro-scale details; Not spatial optimization (BIM)
[43]	Nonlinear Optimization, BIM, Airflow Simulation	High (Healthcare HVAC)	-Hybrid GA-CFD approach Validation: ASHRAE standards	Scalability challenges
[15,18,21]	Nonlinear Optimization, ML, Pattern Identification	High (Predictive control)	-BIM for spatial constraints Method: Phasic Policy Gradient	Limited to static layouts
[44]	BIM, Pattern Identification, ML	Medium (Automation)	-MPC + ML for ventilation DL Model: CNN-LSTM	Slow convergence of GA; Fast predictions come at a cost of degraded accuracy

compares AI-driven HVAC systems.

As shown in Table 1, CFD can simulate scenarios such as disease transmission in healthcare facilities and assess the effectiveness of different airflow management strategies [42,43]. Particularly, CFD equations describe the relationships between external forces and fluid flow and pressure [45], where the 1st-order derivative of a velocity vector is affected by the pressure vector and kinematic viscosity. Table 1 also summarizes the limitations of these AI-driven HVAC systems, including scalability issues, reliance on static layouts, and a lack of real-time adaptability.

• Limited real-time integration: BIM provides static insights and CFD models dynamic airflow, but their real-time operation remains underdeveloped [37,46].
• Data management challenges: Current systems struggle to handle large, dynamic datasets required for responsive indoor air quality and ventilation control [38,39,47].
• Healthcare-specific complexity, such as fluctuating patient loads and activity levels, which demand highly adaptive systems [38,39,40,41,48].
• Computational constraints and a lack of scalable architectures [37,43,44].
• Insufficient validation: Few studies validate these integrated systems in real-world hospital settings, limiting their practical applicability [42].
• Lack of interoperability: Seamless data exchange protocols between BIM, CFD, and ML components are still scarce [15,18,21,23].

The objective of this research is to develop a real-time, adaptive hospital ventilation framework that integrates BIM as a central repository, embeds CFD simulations for dynamic airflow modeling, and leverages ML (including LSTM and PPO) for predictive control and spatial optimization. This approach focuses on the limitations of prior work and enables responsive, multi-objective optimization in complex healthcare environments. This paper addresses these gaps by combining LSTM-based patient flow prediction with CFD-validated PPO control, offering a novel framework for adaptive nonlinear optimization in hospital ventilation.

Even though previous studies have explored AI-driven HVAC control, CFD-based airflow modeling, and BIM integration independently, few have achieved a real-time, closed-loop system that dynamically adapts to occupancy and environmental changes in hospital settings. This study advances the field by combining LSTM-based patient flow prediction, CFD-validated PPO control, and BIM as a central data hub to create a unified framework that optimizes infection control, energy efficiency, and occupant comfort. Unlike prior work that relies on static layouts or offline simulations, this approach enables adaptive, multi-objective optimization in real-world hospital environments.

3. Materials and Methods

3.1. Framework Development

The research design begins with parameter identification, value range determination, and the development of a detailed BIM model for a multi-story hospital building. As illustrated in Figure 1, the proposed BIM-ML-CFD integration framework follows a structured workflow. The BIM model captures architectural geometry, system layouts, and occupancy zones, serving as the foundation for simulation and analysis.

In the research design, patient flow data is processed using LSTM networks to predict peak usage times and inform spatial layout decisions. CFD simulations are conducted using OpenFOAM^® version 12 and ANSYS Fluent^® version 2024 R1, with a structured mesh with refinement near diffusers and patient zones. Industry Foundation Class (IFC) standards are used to facilitate parameter integration and data exchange across platforms. To enhance environmental responsiveness, multiple sensors are deployed to monitor temperature, humidity, and CO₂ levels, with the collected data integrated into the BIM model [37,40]. Machine learning techniques, including PPO and LSTM, are used to train AI agents within a digital twin environment, enabling real-time HVAC control based on predictive and optimization modules. The multi-objective optimization framework balances infection control, energy efficiency, and occupant comfort, guided by HCD principles and validated through case studies. Ultimately, the AI-enabled BIM-ML-CFD system demonstrates potential for generative design automation by comparing its outputs against traditional design methods to evaluate performance and efficiency.

The hospital hosts 700 dental chairs and 200 inpatient dental beds and is a national-level dental center with priorities of public health, building qualities, and operational efficiency, integrating medical care, teaching, research, and preventive care. In the building, the dental clinics are designed to handle a daily outpatient volume of 10,000 patients, while the general hospitalization manages 5000 outpatient and emergency visits per day, making efficient floor layout design essential to optimize patient flow and care delivery.

The hospital BIM model includes comprehensive architectural geometry, such as room dimensions, wall partitions, ceiling heights, and door/window placements, which are essential for accurate airflow and spatial simulations. System layouts include mechanical (HVAC ductwork, air handling units, exhaust systems), electrical (lighting, power distribution), piping (water supply, drainage, medical gas, lab gas and waste, steam and condensate, fire protection, and wastewater and biohazard disposal), and security (access control and surveillance zones). These systems are modeled to reflect real-world installation constraints and operational interdependencies. Occupancy zones are defined based on functional use (e.g., treatment rooms, waiting areas, corridors), patient flow patterns, and staff movement data, enabling dynamic simulation of space utilization and congestion. Material properties (e.g., thermal conductivity, surface roughness, and permeability) are assigned to architectural elements to support CFD and energy modeling. The BIM model was developed using Autodesk Revit for geometry and system modeling, and Navisworks for coordination, clash detection, and integration with simulation tools. Data from facility design documents and stakeholder input were incorporated to ensure fidelity to the actual hospital layout. The model serves as the foundation for AI-driven analysis, CFD simulations, and optimization processes throughout the study. Additionally, this research uses both simulation data and real-world data from application scenarios of complex medical processes. The assumption is that nonlinear building dynamics have complex airflow patterns and require precise modeling.

3.2. Energy Calculation

Equation (1) is a generalized Momentum–Energy Transport Equation using the material derivative ${\displaystyle \frac{D}{Dt}}={\displaystyle \frac{\partial }{\partial t}}+u·\nabla$. It retains viscous dissipation (Φ) from mechanical energy loss and heat conduction (k∇²T) and assumes ideal gas behavior. Included in the equation, ($c_v$) represents the specific heat at constant volume, and ($c_p$) represents the specific heat at constant pressure. The assumptions include Newtonian fluid, Fourier’s law for heat conduction, negligible radiative heat transfer, and low Mach number (neglecting compressibility effects in energy terms). In the equation, $\rho$ is the fluid density, $u$ is the velocity vector (u, v, w), t is time, $p$ is pressure, $\mu$ is dynamic viscosity, and ${\nabla }^{2}u$ is the viscous term of the Laplacian of velocity. The $\rho {\displaystyle \frac{du}{dt}}$ is unsteady inertia, representing the rate of change in momentum, $\rho c_p{\displaystyle \frac{DT}{Dt}}$ is advection, describing the transport of momentum by the fluid itself, $-\nabla p$ is the pressure gradient or driving force, and $\mu {\nabla }^{2}u$ is viscous forces and accounts for internal friction in the fluid. For incompressible flows, $\left(u·\nabla \right)=0.$

$$\rho {\displaystyle \frac{Du}{Dt}}+\rho c_p{\displaystyle \frac{DT}{Dt}}=-\nabla p+\mu {\nabla }^{2}u+k{\nabla }^{2}T+\mathsf{\Phi },$$

(1)

Equation (2) explains mass–energy continuity. In the equation, the thermal expansion coefficient $\beta =-{\displaystyle \frac{1}{\rho }}\left({\displaystyle \frac{\partial \rho }{\partial T}}\right)$. The $\rho \beta {\displaystyle \frac{DT}{Dt}}$ is for incompressible flows with Boussinesq approximation. Equation (1) solves airflow–pathogen transport with thermal coupling, and Equation (2) ensures mass–energy consistency in negative-pressure zones.

$${\displaystyle \frac{D\rho }{Dt}}+\rho \nabla ·u+\rho \beta {\displaystyle \frac{DT}{Dt}}=0,$$

(2)

3.3. Reinforcement Learning

The ML and IoT control module of the framework leverages Kaggle Healthcare Datasets [49], including patient visits, hospital utilization, and related healthcare statistics, to implement RL. Figure 2 shows how the PPO algorithm predicts and upgrades HVAC controls. The positive reward functions are for better thermal comfort, minimized energy consumption, and reduced pollutant levels in the breathing zones. On the other hand, the negative reward functions are for the overuse of energy, poor airflow distribution, or high pollutant levels near occupants. Appendix A shows the pseudo-code of the algorithm.

3.4. CFD Simulation Setup

The key steps of the PPO workflow include the state space, action space, advantage-reward space, clipping threshold, surrogate loss, and integration with CFD. Following the technical framework outlined in [23], the CFD setup employed a structured mesh strategy with localized refinement near supply diffusers and patient zones to capture detailed airflow behavior. The computational grid consisted of approximately 5.2 million hexahedral elements, with minimum and maximum layer thicknesses of 0.01 m and 0.3 m, respectively. An expansion ratio of 1.1 was used to ensure smooth gradation between mesh layers, minimizing aspect ratio distortion and supporting solver convergence. The simulations were conducted using a pressure-based steady-state solver with the SIMPLE algorithm for pressure–velocity coupling. Turbulence was modeled using the Renormalization Group (RNG) k–ε model, which is well-suited for indoor airflow scenarios involving recirculation and moderate turbulence. Boundary conditions included a power-law velocity profile and turbulent intensity at the inflow, fixed static pressure at the outflow, and symmetrical boundaries at the top and lateral surfaces. Building walls and ground surfaces were treated using a power-law formulation for smooth surfaces and standard wall functions. This configuration provides a robust and reproducible simulation environment for evaluating airflow and aerosol dispersion in clinical settings.

The state space captures BIM- and CFD-based parameters like airflow velocity, pressure, and pollutant concentration in the breathing zone. Real-time sensor feedback refines the state. The action space includes HVAC control actions such as changing air inlet temperature, adjusting vent angles, or modifying fan speeds to regulate ventilation. In the advantage estimation, PPO relies on Generalized Advantage Estimation (GAE) for stable and efficient updates, which are deployed using reward functions. The clipping threshold (ε) ensures updates do not deviate excessively, improving stability. The surrogate loss function calibrates the policy and maintains the exploration–exploitation balance. At each step, actions taken by the RL agent modify parameters in the CFD simulation, providing realistic feedback for training.

3.5. Hypothesis and Scenario Selection

This study hypothesizes that the proposed BIM-ML-CFD framework, trained using reinforcement learning, can significantly outperform traditional static HVAC systems in reducing airborne pollutant exposure and energy consumption. This hypothesis directly supports the study’s aim of developing a dynamic, multi-objective optimization system for hospital ventilation that integrates real-time environmental data, spatial modeling, and adaptive control.

Equation (3) combines multiple objectives into a single objective by summing each objective, multiplied by a user-defined weight. The weight of each objective reflects its relative importance. It shows the weighted sum method for the Multi-Objective Optimization Problems (MOOPs) of this study, subject to $g_j\left(x\right)\ge 0,j=\mathrm{1,2},\dots ,J$, $h_k\left(x\right)=0,k=\mathrm{1,2},\dots ,K$, $x_i^{\left(L\right)}\le x_i\le x_i^{\left(U\right)},i=\mathrm{1,2},\dots ,n$

$$Minimize:F\left(x\right)={\sum }_{m=1}^M{w}_m{f}_m\left(x\right)$$

(3)

This study investigates four organizational scenarios derived from the design, optimization, and construction phases of a real-world dental hospital project. A dental hospital was selected because operation units in such facilities exhibit higher pollutant concentrations in the breathing zone from aerosol-generating procedures, making them ideal for testing adaptive and dynamic indoor air quality (IAQ) control strategies. The framework’s flexibility is particularly valuable in these high-risk environments. The four scenarios are structured to reflect the progressive development of the hospital’s spatial and mechanical systems:

• Floor layout optimization is performed using LSTM-based patient flow prediction, which leverages historical check-in and appointment data to anticipate congestion and inform spatial arrangement decisions.
• AI-driven spatial analysis and optimization integrate BIM geometry with reinforcement learning to evaluate airflow patterns, thermal comfort, and contaminant dispersion, ensuring that design choices align with both clinical and environmental performance goals.
• Real-time adaptive HVAC system modeling simulates dynamic responses to occupancy and contaminant levels, enabling the system to adjust ventilation rates and pressure zones based on predicted usage patterns.
• Multi-objective optimization balances competing priorities—such as energy efficiency, infection control, and patient comfort—using a weighted scoring system informed by ASHRAE and WHO indoor air quality guidelines.

Each scenario builds upon the previous one, forming a comprehensive and iterative design process that reflects real-world decision-making in healthcare facility planning. The design and planning process also includes stakeholder input, spatial constraints, and infection control priorities. This structured approach ensures that the framework is technically robust and practically applicable across different phases of hospital development.

4. Results

4.1. Case Study 1: Floor Layout Optimization Using LSTM-Based Patient Flow Predictions

Figure 3 shows how BIM models were developed for simulation and analysis purposes in a hospital project. The scenarios in Figure 3 were developed through an iterative design process informed by BIM-based spatial modeling, AI-driven patient flow prediction using LSTM networks, and stakeholder input. Each layout was evaluated based on key performance indicators such as flow line clarity, circulation length, proximity of functional zones, and staff working conditions. Figure 3a presents a shared-function layout. Figure 3c–e explore surrounding and sequential arrangements with varying degrees of flow efficiency and environmental quality.

Figure 3b was selected as the optimal design strategy due to its parallel layout, which ensures non-intersecting doctor-patient flow lines, short circulation paths, and improved staff working conditions. This scenario demonstrated superior performance in BIM simulations and AI-based optimization, particularly in reducing congestion and enhancing air quality in high-traffic zones. The scenario development process is essential to the study as it provides a structured framework for evaluating spatial configurations and their impact on IAQ, patient comfort, and operational efficiency. These insights directly inform the ventilation optimization strategies and support the broader applicability of the proposed framework to other healthcare environments.

Figure 3 utilizes simulation and real-world data to evaluate the impact of nonlinear building dynamics on airflow and spatial efficiency in hospital environments with various layout strategies. Insights from the figure are provided below.

• Designs that maintain clear separation between doctor-patient flow lines and short circulation paths (e.g., Figure 3b) are shown to improve operational efficiency and reduce cross-contamination risks.
• Impact of Room Proximity: Layouts where medical rooms are placed close to waiting areas (Figure 3c,e) result in suboptimal working environments, potentially increasing exposure to airborne pathogens and reducing staff comfort.
• Design Selection Justification: Among the evaluated layouts, Figure 3b was selected for implementation due to its parallel arrangement, which optimizes spatial flow, minimizes congestion, and supports better ventilation control.
• Human-Centric Spatial Planning: The analysis highlights how architectural layout directly influences HVAC performance, patient experience, and staff workflow, reinforcing the need for integrated design strategies in healthcare infrastructure.

Figure 4 shows a floor plan layout for the AI-enabled BIM simulations to improve the dental medical center design strategy. This plan adopts a model divided into four areas (Common Inspection Area, Patient Waiting Area, Medical Treatment Area or MDT, and Internal Medical Area). The pattern analysis of the BIM simulation analysis shows that the patient and medical care areas are divided and do not interfere with each other. The upper area is the Common Inspection Area and the public transportation area. The Patient Waiting Area is between the Common Inspection Area and the MDT. The middle area is the MDTs, and the relatively independent area at the bottom is the Internal Medical Area. The layout of the entire floor follows the identified and selected design strategy (Figure 3b).

Building on insights from historical data, the floor layout incorporates principles of Human-Centered Design (HCD), focusing on three key areas: (1) Patient-centered spatial design; (2) personalized spatial configurations; and (3) medical-centered space planning. Each floor is equipped with advanced diagnostic facilities to enhance convenience, including panoramic imaging, dental X-rays, cone-beam-computed tomography (CBCT), and ultrasound, minimizing the need for vertical movement and reducing stair traffic. Each floor features a dedicated patient service area, and a sky café on the top floor provides a tranquil resting space for patients and visitors (labeled “A” in the figure). Leisure seating is thoughtfully positioned near waiting areas adjacent to the atriums with ample daylight, helping alleviate patient anxiety during treatment (labeled “B” in the figure). Bathrooms are strategically placed within a maximum service radius of 60 m, with additional third-gender facilities to accommodate diverse needs (labeled “C” in the figure).

The design of inner courtyards ensures that nearly every consulting room benefits from natural light and ventilation (labeled “D” in the figure). For personalized spatial layouts, treatment chairs and facilities in different dental departments are tailored to their specific functional requirements. In specialized care areas, every two consulting rooms share a private courtyard to enhance patient comfort. To address infection control during epidemics, most consulting rooms are configured as single-occupancy spaces. The medical-centered design prioritizes staff well-being by integrating rest areas on each floor and ensuring that office spaces receive abundant natural light, creating a well-designed working environment for healthcare professionals.

4.2. Case Study 2: AI-Driven Spatial Analysis and Optimization

One of the key concerns in this project was evaluating whether the spatial arrangements were adaptive and optimized for hospital functions. Figure 5 illustrates the integration of AI-enabled BIM simulation and performance enhancement tools, which were employed to improve design efficiency and operational outcomes. Specifically, patient flow simulations were conducted to refine spatial layouts and support a patient-centered design approach.

Figure 5a depicts the patient’s check-in stage within a dental treatment unit, where the patient (P) enters when the doctor (D) is already present, and the nurse (N) arrives through a separate entrance. Figure 5b presents a 3D visualization generated from the BIM model for this scenario. Figure 5c highlights the inclusion of an extraoral dental suction unit (EAS), designed to capture aerosols and reduce airborne contaminants. Figure 5d shows the actual equipment arrangement within the treatment unit.

Figure 5e,f provide top and 3D views of the activity area designated for treatment, discussion, and preparation. Similarly, Figure 5g,h illustrate the spatial configuration of the treatment process area. Figure 5i,j detail the layout of key devices, including the EAS, ground junction box (J), computer (C), water faucet and sink (W), and microscope.

To further enhance operational efficiency, the BIM-ML-CFD system was implemented to predict peak usage times and recommend adjustments to mitigate congestion (e.g., in waiting and diagnostic areas). By integrating the AI-driven BIM simulations with Kaggle Healthcare Datasets [49] and historical patient flow data (e.g., check-ins, appointments, walk-ins), LSTM networks were utilized to analyze movement patterns, identify bottlenecks, and streamline traffic flow. The selection of EAS devices for each treatment unit was also supported by simulation results, recognizing that airborne contaminant requirements vary during a pandemic. The use of portable EAS units provided the necessary flexibility to adapt to evolving clinical and environmental conditions. Based on these simulations, this AI-BIM model recommended installing one EAS device and two ground junction boxes per unit to improve air quality during procedures, with placements illustrated in Figure 5i,j. Three representative scenarios (Figure 5a,e,g) were selected for CFD analysis and further design refinement, as detailed in Case Studies 3 and 4.

4.3. Case Study 3: Real-Time Adaptive HVAC System

Next, the BIM-ML-CFD system was applied to calibrate the partition board height. A BIM model was developed, and STAR-CCM+ version 2406 software analyzed airflow in the treatment unit (see Figure 6a). Figure 6b also displays a speed cloud chart from BIM-CFD simulations, which guided the air conditioning system, negative pressure zones, and partition layout. This ensured isolated airflow between adjacent spaces, preventing cross-contamination. Figure 6c,d show airflow streamlines from patient breathing and velocity vectors, respectively, providing single-parameter insights into aerosol dynamics. Figure 6e shows CFD airflow streamlines for the modified wall heights, which were increased from the traditional 1.5 m to 1.8 m to ensure independent airflow separation between adjacent rooms. This decision was supported by using the simulations on 1.5, 1.6, 1.7, and 1.8-m partition boards and observing the corresponding airflow streamlines. When air from the upper distribution system flows downward, it first passes the patient’s head and bed before splitting into two paths: one portion exits through the corridor return air system, while another flows along the wall before returning to the corridor, effectively preventing aerosol cross-contamination. Figure 6f presents the CFD velocity cloud map used to adjust the air conditioning parameters. During treatment, the airflow supply (peaking at 1 m/s) effectively captures oral bacteria from the patient’s breathing zone and removes them through the return system, significantly reducing bacterial circulation time in the room. Figure 6g displays the velocity vector map without an EAS device, revealing a recirculation vortex on the patient’s right side that traps exhaled pathogens. In contrast, Figure 6h demonstrates the improved airflow with an EAS device installed, where over 85% of the airflow is directed to the suction system, ensuring rapid aerosol removal and maintaining ultra-low contamination levels.

4.4. Case Study 4: Multi-Objective Optimization

Figure 7 summarizes the parametric analysis, showing that 80% of cell velocities range between 0.00021726 and 0.38049 m/s, with a space-averaged velocity of 0.19 m/s. Based on Equations (2) and (3), this optimized configuration achieves 51.85% energy savings compared to conventional systems operating at 1.9016 m/s.

Table 2. Summary of BIM-ML-CFD simulation.

Pressure at the Return Vent (Bar)	Speed at Supply Diffuser (m/s)	Aerosol Existence Time (s)
0	1	351
−0.5	1	159
−1	1	67
−2	1	36
0	2	418
−0.5	2	182
−1	2	105
−2	2	81

lists the air pressure, airflow speed, and time of aerosol generated from a patient’s mouth when they stay in the diagnosis and treatment space under different working conditions. By using discrete multiphase flow (DMP) analysis, liquid particles (a.k.a. aerosols) are released from the mouth, and the time from the start of liquid release to the final movement to the return air outlet is counted to calculate and improve the air supply outlet wind speed (or flow rate) and the return air outlet negative pressure. Based on simulation analysis, the residence time of droplet particles under several different combinations is counted. As shown in the table, when the negative pressure is [0, −0.5] Bar at the return vent and the wind speed is 1 m/s, the aerosol existence time is in the range of [351, 159] seconds. Meanwhile, it is also found that under the same negative pressure, the higher the wind speed, the longer the droplet residence time. The reason is that when the wind speed is higher, the vortex generated inside the room is more obvious, so the air supply speed should not be too large. The HVAC setting of negative pressure (−0.5 Bar at the return vent) and wind speed (1 m/s) can reduce the aerosol existence time from 418 s to 159 s, for a 1.6 times improvement in air quality and a 61.96% reduction in pollutant exposure.

Table 2 indicates that the derivative values increase as pressure approaches 0, and aerosol persistence is highly sensitive to small changes in pressure near atmospheric levels. At 2 m/s, the system is more responsive to pressure changes, especially near 0 Bar, suggesting stronger airflow effects on aerosol clearance. The second-order derivatives of Aerosol Existence Time with respect to Pressure at the Return Vent for the supply diffuser speeds are as follows: Pressure Interval [−2, −1], Speed 1 m/s (s/Bar²) = 153.0, and Speed 2 m/s (s/Bar²) = 130.0; Pressure Interval [−1, −0.5], Speed 1 m/s (s/Bar²) = 400.0, and Speed 2 m/s (s/Bar²) = 636.0. Following the same mathematical structures as Equations (1) and (2), the infection control efficiency is related to the second-order derivatives and falls within the range of [17.69%, 37.11%] in medical treatment units.

Figure 7 illustrates how key hospital design objectives, such as maintaining distinct circulation paths for patients and medical staff, optimizing spatial efficiency, and supporting ergonomic working conditions, are achieved through the spatial strategies implemented in the Consultation Room layout (Figure 3b) and the multidisciplinary team (MDT) zones shown in Figure 4 of Case Study 1. These spatial configurations were developed based on direct user feedback and collaborative discussions with stakeholder groups, ensuring alignment with human-centered design (HCD) principles, as further detailed in Figure 5 of Case Study 2. To enhance infection control and resource efficiency, the partition boards surrounding each treatment unit were strategically modified to reduce material usage, which improved aerosol containment, as demonstrated in Figure 6 of Case Study 3. Throughout the design process, occupant comfort was prioritized by adhering to established standards and guidelines related to thermal comfort, spatial accessibility, and indoor air quality. Hence, using equal weights in the MOOP defined in Equation (3), this case study demonstrates that the key functions of infection control, energy efficiency, and occupant safety and comfort have been successfully achieved in the hospital environment.

Unlike prior offline CFD optimizations [16], which simulate airflow for fixed layouts and occupancy scenarios, this framework dynamically updates boundary conditions using real-time BIM and IoT data. Where traditional RL-HVAC studies [16] train policies on historical datasets, the PPO agent interacts with a CFD-informed environment model that evolves with LSTM-predicted patient flows (Figure 6). This enables continuous adaptation to transient occupancy patterns, especially for hospitals with highly variable aerosol generation. For example, compared to Wang et al. [19] who modified partition heights using genetic algorithms and static CFD, this system adjusts HVAC settings (e.g., −1 Bar pressure at the return vent, 1 m/s airflow) in response to simulated aerosol dispersion (Table 2 and a 61.96% reduction in exposure time), which optimizes both occupant comfort and infection control. This integration of real-time data, predictive modeling, and CFD validation represents a paradigm shift from offline design to operational intelligence.

5. Discussion

This study presents transformative advancements in hospital ventilation design through an intelligent closed-loop BIM-ML-CFD framework integrated with AI-driven Digital Twin technologies. This system represents a paradigm shift from conventional static approaches by implementing real-time, physics-informed airflow control that dynamically adjusts HVAC parameters in response to changing environmental conditions and space utilization patterns. The experimental results demonstrate unprecedented performance: 1.6 times better air quality and a 61.96% reduction in airborne pathogen spread (Table 2), substantially exceeding the capabilities of previous RL-based HVAC systems [21] and offline CFD approaches [16,17] while maintaining thermal comfort.

Furthermore, the system effectively balances thermal comfort with holistic considerations, including temperature, energy consumption, and pollutant levels. Experiment results highlight significant energy savings, achieving a 51.85% improvement in airflow velocity changes and air quality optimization, supported by reward functions designed for performance enhancement. This is compared to the MPC approach in [20], which developed building ventilation systems with a focus on air quality, energy cost, and CO₂ emissions. Specifically, the system improved air quality (e.g., CO₂ concentration reduction) and achieved energy cost savings. The BIM-ML-CFD method and the MPC approach implemented dynamic airflow adjustment and management based on real-time data and optimization objectives. The BIM-ML-CFD method was verified by the experiment in a hospital, and the MPC approach was for a university building; hence, infection risks, occupant comfort, and ergonomics considerations (e.g., dentists’ or medical personnel’s postures, patients’ chairs, and examination beds’ locations) are not included as the primary variables of the latter study.

Key contributions of this research include the development of a dynamic BIM-ML-CFD framework that integrates real-time sensor data, data-driven analysis, and reinforcement learning through the PPO algorithm. The framework employs time-series analysis using LSTM networks and Kaggle Healthcare Datasets to predict patient flow, identify bottlenecks, and propose adjustments to mitigate congestion in critical zones. Simulations revealed effective strategies for spatial optimization and aerosol management, reducing cross-infection risks in treatment areas. The system reduces airborne pathogen exposure by 61.96% (159 s vs. 418 s residence time). Chen et al. [31] identified occupancy as a dynamic input in their LSTM-based thermal load prediction framework, significantly enhancing model accuracy. Occupancy data were obtained from historical building operation records or simulated using EnergyPlus, a physics-based modeling tool, enabling the representation of realistic daily occupancy patterns and fluctuations. This approach allowed the model to more accurately capture peak thermal loads, particularly during working hours, reducing overall prediction error to below 5%. In contrast, this research placed greater emphasis on building operation scenarios, categorizing them specifically for ergonomic analysis and HCD considerations. Specific interventions, such as partition board height adjustments and negative pressure systems, were designed based on CFD simulations to ensure independent airflow lines between adjacent spaces.

This research also applied advanced computational methodologies and HCD principles to optimize the layout and operational efficiency of a large-scale hospital project. These principles enhanced accessibility, functionality, and patient comfort. By integrating time-series analysis and simulation tools, the study enabled dynamic predictions of patient demand, resource optimization, and improved service delivery. These outcomes underscore the effectiveness of combining design innovation with advanced technologies to enhance healthcare infrastructure and patient care.

The dynamic BIM-ML-CFD framework developed in this study incorporates real-time sensor data and machine learning to enable dynamic HVAC control for infection control, operational efficiency, and energy optimization. The integration of reinforcement learning and CFD simulations introduced a patient-centric model for hospital design that prioritizes comfort, accessibility, and functionality. These innovations provide actionable insights into optimizing healthcare environments through scalable, data-driven solutions. Moreover, the combination of simulation-based resource planning and predictive modeling establishes a new paradigm for designing adaptable and strategically aligned healthcare facilities.

This research sets a benchmark for future healthcare projects by bridging the gap between technological advancements and practical applications in public health infrastructure. For instance, Wang et al. [19] utilized Euclidean distance, Dice coefficient, and a force-directed graph algorithm for layout selection and optimization, incorporating the Input-Controlled Spatial Attention U-Net model for precise segmentation of building regions. In contrast, the BIM-ML-CFD framework in this study integrates machine learning with computational fluid dynamics for real-time environmental factors such as airflow and energy efficiency, alongside layout optimization. By combining reinforcement learning and dynamic simulations, the proposed approach delivers a more holistic and adaptive solution to hospital design challenges.

This study emphasizes the necessity of a comprehensive vision for modern healthcare infrastructure that integrates scale, quality, innovation, and strategic foresight [1,2,4,9,10,11,12]. It sets a foundation for multidisciplinary collaboration and fosters resilience and preparedness in tackling global healthcare challenges. By adopting cutting-edge technologies, including PPO algorithms for patient flow optimization and CFD for modeling airflow and ventilation, this hospital project exemplifies precision and sustainability. Ultimately, it establishes a benchmark for current and future healthcare demands with innovative, scalable, and effective solutions.

This work establishes a new paradigm for healthcare infrastructure, opening several research avenues, including the clinical validation of recovery outcomes, generalization to non-hospital settings, and the integration of waste heat recovery systems. For example, the responsive feedback loop of the framework enables dynamically adaptive infrastructures for routine operations and emergency scenarios, which is a capability previously absent in prior RL/CFD studies.

HCD principles can be integrated into the BIM-ML-CFD framework by incorporating patient comfort metrics, such as Predicted Mean Vote (PMV), thermal sensation, and perceived congestion levels, as dynamic feedback signals in an RL process [50]. These metrics can be derived from both sensor data (e.g., temperature, humidity, and CO₂ concentration) and human movement patterns captured via simulations and historical flow data. In this framework, the RL agent continuously evaluates ventilation strategies for aerosol containment and energy efficiency and their impact on patient comfort. In this case, PMV values can be calculated for different zones and used to respond to observed thermal discomfort. Similarly, congestion levels are estimated from real-time occupancy and flow predictions to help avoid psychological stress or dissatisfaction. This HCD integration ensures that ventilation optimization is not solely driven by mechanical performance but also by experiential quality. By embedding these comfort-related parameters into the reward function, the BIM-ML-CFD system learns to balance infection control with environmental comfort, resulting in more holistic and adaptive design outcomes.

Scalability of the AI-driven BIM-ML-CFD framework should be considered for deployment in large hospitals or multi-building medical campuses. As facility size or capacity increases, challenges such as decentralized HVAC systems, variable occupancy patterns, and data heterogeneity become more pronounced. A possible solution is a modular and hierarchical architecture in which each building or zone operates a localized optimization module for real-time data processing and decision-making, coordinated by a central supervisory module to ensure consistency and interoperability. Innovative technologies such as cloud-based data pipelines and federated learning techniques can be employed to manage large-scale data integration, which preserve privacy and reduce bandwidth demands [51]. Furthermore, scalability-aware RL algorithms can be incorporated to dynamically adjust parameters [52]. This distributed and adaptive design enables incremental deployment and integration with existing systems.

6. Conclusions

This study has advanced hospital ventilation systems through an integrated BIM-ML-CFD framework that dynamically improves thermal comfort, energy efficiency, and infection control. The results demonstrate substantial improvements over conventional approaches, including a 61.96% reduction in aerosol exposure time and 51.85% energy savings. By combining real-time IoT monitoring with predictive LSTM networks and CFD-validated reinforcement learning, the system achieves responsive environmental control that adapts to changing clinical demands. Compared with static ventilation designs, this approach offers hospitals a robust solution for balancing competing operational priorities.

The framework has measurable benefits across multiple domains. In clinical settings, enhanced airflow patterns and pressure differentials effectively mitigate the transmission of airborne pathogens, directly protecting patients and staff from infections. The system’s energy-efficient operation reduces hospital operational costs and meets stringent environmental standards, demonstrating that sustainability and patient safety can be mutually achievable goals. Furthermore, the adaptable design ensures resilience against emerging public health challenges, providing a future-proof foundation for hospital infrastructure. For the evolution of hospital HVAC design standards, particularly in the context of airborne infection control, the findings help quantify airflow separation, aerosol containment, and thermal comfort. Based on current ASHRAE guidelines (e.g., Standard 170 for ventilation in healthcare facilities) and WHO recommendations on indoor air quality, the integration of AI-driven optimization and HCD metrics could inform future revisions of these standards, promoting more responsive, data-driven, and patient-focused ventilation strategies in healthcare environments.

Future research should focus on developing adaptive HVAC frameworks that integrate advanced computational models, real-time data, and emerging technologies. Key areas include leveraging AI and IoT for predictive maintenance and real-time adjustments, exploring multi-objective optimization techniques for balancing thermal comfort, energy efficiency, and cost, and evaluating health outcomes like recovery times. Environmental impact modeling could aim for carbon neutrality through renewable energy and sustainable materials. Scalability and resilience testing in diverse hospital settings are crucial for handling challenges like pandemics. These efforts will enhance HVAC performance, bridging engineering innovation with practical, measurable societal and environmental benefits.