A Data-Driven Multi-Objective Optimization Approach for Improving Energy and Comfort Performance in Residential Buildings

Abstract

Examining energy-related issues and comfort conditions in residential buildings is essential for achieving sustainable development. Accordingly, this study aims to optimize building envelope design parameters—including air-tightness values derived from empirical measurements—to strike a balance between energy consumption, thermal comfort, and visual comfort. A multi-stage, systematic methodology is proposed, featuring a hybrid artificial intelligence-based framework for prediction and optimization, grounded in building performance simulations. A BlowerDoor test was conducted on the residential building under study, and the results were integrated into a building performance model. Data generated through these simulations were used to compare the performance of various machine learning meta-models, including Artificial Neural Networks, Support Vector Regression (SVR), and XGBoost. The SVR model, which demonstrated the lowest error rate and highest reliability across most objectives, was selected as the meta-model for the optimization phase. This model served as the objective function for a multi-objective optimization process carried out using the Non-dominated Sorting Genetic Algorithm II (NSGA-II). The results of this surrogate-model-based optimization indicate a 21.69% reduction in energy consumption, alongside improvements of 17.2% in visual comfort and 20% in thermal comfort compared to the baseline state. Furthermore, the analysis reveals that air-tightness influences daylighting performance through its interaction with other parameters, such as the Window-to-Wall Ratio (WWR) and building orientation. Finally, the SHapley Additive exPlanations (SHAP) method was employed to determine the relative importance of the design parameters following the multi-objective optimization. The analysis results indicate that the order of influence of various parameters differs depending on the specific objective. While the window-to-wall ratio (WWR) and orientation are the primary factors for enhancing visual comfort, factors such as orientation, infiltration, and the thermal transmittance of building components are the most significant for improving energy efficiency and thermal comfort. Consequently, the integration of data-driven approaches—based on surrogate models—into building performance assessment can contribute significantly to the sustainability of residential buildings and offer a new research framework for future studies.

1. Introduction

The building sector accounts for a significant share of total energy consumption. Current regulatory standards mandate the design of buildings that are as energy-efficient as possible [1]. Energy demand in buildings has become a major concern within the construction sector, particularly in recent years. Europe aims to achieve climate neutrality by 2050. Energy production and use account for more than 75% of greenhouse gas (GHG) emissions in the European Union. Furthermore, the European Energy Performance of Buildings Directive (EPBD) aims to achieve a highly efficient and decarbonized building stock, recognizing that approximately 50% of final energy consumption is used for heating and cooling [2].

High-efficiency buildings feature well-insulated envelopes, minimized thermal bridges, and improved construction quality. High airtightness, in particular, is one of the most critical physical characteristics of high-performance buildings. Consequently, factors such as airtightness are becoming increasingly important regarding energy demand [3]. In the 1970s, as energy efficiency gained prominence, various measures were implemented to reduce air infiltration and leakage. Many countries established limits for airtightness values within their national regulations. The most widely used parameter for measuring airtightness is the air change rate per hour at a pressure difference of 50 Pa between the indoor and outdoor environments; this parameter is typically determined based on factors such as the ventilation strategy or building type [4]. Buildings are not highly airtight, a situation that creates challenges regarding energy efficiency and indoor air quality. Air infiltration is the entry of outdoor air through cracks in the building envelope or other openings, such as open windows and exterior doors [5]. If air tightness—or conversely, air leakage—is not properly controlled, it leads to high energy consumption for space heating and cooling [6,7].

Previous studies on air leakage have largely focused on buildings with standard layouts, such as residences, offices, and classrooms [8]. Literature indicates that air tightness has the most significant impact on energy consumption in cold climate regions. Its high impact in cold climates reflects the fact that heating energy consumption is relatively higher than cooling energy consumption. Furthermore, previous studies have estimated the energy impact of leakage on heating demand in temperate climate regions to be between 2 and 20 kWh/(m²·yr) [9]. Poza–Casado et al. [9] investigated the impact of uncontrolled airflow through the building envelope on the energy consumption of residential buildings in Spain. They determined that leakage had an energy impact ranging from 2.43 to 19.07 kWh/(m²·yr) regarding heating demand, whereas its impact on cooling demand was less significant.

Various simulation methods are employed to estimate the impact of air leakage on building energy consumption. Bhandari et al. [10] developed an approach based on a simple regression model for hourly leakage calculations. It was observed that when the proposed method was used, the energy savings rate for retrofitted buildings fell within 5% of the estimates obtained from an online calculator. Vera Alves et al. [3] compared four different approaches regarding air tightness using the TRNSYS energy simulation tool. The difference in estimated energy consumption among the models was found to range from 6.1% to 12.2%. In another study, Ren et al. [11] developed three different infiltration models to evaluate improvements in air infiltration modeling for residential energy efficiency. The study, which assessed eight residences, reveals differences of up to 37% in calculated average infiltration rates and up to 13% in modeled heating and cooling energy demand among these three models. Ren et al. [12] developed an integrated airflow model for AccuRate, a residential energy efficiency rating tool. The model proposed in the study confirms that infiltration rates in 10 houses could be reduced by an average of 52% by sealing major air leakage paths, resulting in an average 7% reduction in heating and cooling loads. Liu et al. [13] analyzed the impact of increased building airtightness on energy consumption and indoor temperature. Simulation results indicate that as the airtightness level improves during summer months, indoor temperature rises, while comfort levels decrease and air conditioning loads increase. These results confirm that air tightness significantly affects not only energy consumption [14] but also changes in indoor air quality [15,16] and occupant comfort [17,18].

Since the airtightness of the building envelope plays a significant role in indoor air quality (IAQ), it can help reduce the infiltration of outdoor pollutants such as PM2.5 [19,20], as well as mitigate CO₂ accumulation and thermal discomfort in spaces lacking adequate mechanical ventilation [21]. Exposure to various air pollutants can pose serious health risks to building occupants. Furthermore, poor indoor air quality leads to respiratory symptoms, asthma exacerbations, and issues regarding occupant focus and attention [22]. In this context, the airtightness of the building envelope plays a pivotal role in determining indoor environmental conditions. Overall, ensuring the airtightness of the building envelope is crucial for occupant health and comfort, the effective performance of ventilation systems, and building energy efficiency.

In recent years, multi-objective optimization approaches—which evaluate the combined effects of building envelope design decisions on energy consumption, thermal comfort, and visual comfort—have gained increasing importance. In this context, the integration of machine learning-based surrogate models with simulation tools and optimization algorithms enables the generation of a larger number of design alternatives at a lower computational cost. Wu et al. [23] investigated the impacts of window and shading systems in office buildings on energy consumption, as well as visual and thermal comfort performance, using an evolutionary-based multi-objective optimization method. To quantify the influence of design parameters on optimization objectives, they combined machine learning models with SHapley Additive exPlanations (SHAP) analysis. The proposed method significantly improved energy consumption, thermal performance, and visual performance. Duan et al. [24] proposed a machine learning-based multi-objective optimization framework for the integrated design of windows and photovoltaic (PV) louvers in office buildings located in cold climates. They utilized a machine learning-based multi-objective optimization approach to optimize daylighting, net energy use intensity, and the percentage of thermal discomfort. Through SHapley Additive exPlanations (SHAP) analysis, they identified the relative importance of design parameters with respect to these performance metrics. Hosamo et al. [25] proposed a system that integrates Building Information Modeling (BIM), machine learning, and the NSGA-II algorithm to examine the impact of building envelope design parameters on energy consumption and to determine an optimal design. In another study, Duan et al. [26] presented a data-driven framework—based on XGBoost, SHAP, and NSGA-II—to examine the relationship between building envelope design parameters and performance objectives, specifically energy, thermal comfort, and daylighting, in residential buildings. In the multi-objective optimization conducted as part of this study, economic criteria were also evaluated alongside the other objectives.

The reviewed studies demonstrate the application of data-driven optimization approaches to address challenges related to building performance. However, existing research primarily focuses on design variables such as window geometry, facade-integrated shading elements, or the thermal properties of the building envelope. The integration of on-site measured air-tightness data with other building envelope parameters—and the subsequent analysis of its impact on energy consumption, thermal comfort, and visual comfort within a unified optimization framework—remains an under-researched area. The primary objective of this study is to develop a data-driven, multi-objective optimization framework that integrates on-site measured air-tightness data with other building envelope design parameters to balance energy consumption, thermal comfort, and visual comfort performance. To enhance the reliability of the model, a Blower Door Test (BDT) was conducted on a north-oriented residential building in Balıkesir, and the results were integrated into the building performance model. The methodology employed includes three distinct machine learning models, a surrogate model based on the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) metaheuristic, and SHAP analysis to evaluate energy efficiency, thermal comfort, and visual comfort.

The study is structured as follows: Section 1 provides an overview of the context and background of the research. Section 2 details the general framework and implementation steps of the integrated methodology proposed in the study. Section 3 presents a comprehensive analysis and discussion of the findings. Section 4 summarizes the overall conclusions and offers recommendations for future research.

2. Materials and Methods

This study developed a multi-stage, systematic method to optimize the energy performance, as well as the visual and thermal comfort, of an existing residential building (Figure 1). The research focuses on a north-facing, single-occupancy residence located in Balıkesir. The first stage of the study involved determining the geometric and structural characteristics of the residence. In the second stage, a Blower Door test was conducted to assess the residence’s existing airtightness and to incorporate the results into a parametric energy model. Based on the Blower Door test, the airtightness performance was quantified as the air exchange rate (ACH50)—expressed in air changes per hour (h⁻¹)—at a pressure difference of 50 Pascals. Parametric modeling of the residence was performed using Rhino/Grasshopper. This model was integrated into a Python environment, and building performance simulations were conducted based on the measured airtightness. To reduce the computational cost of optimization, an SVM-based surrogate model was created, and a Multi-Objective Optimization (MOO) process was executed. Furthermore, the relative impact of each parameter on energy consumption, thermal comfort, and visual comfort objectives was evaluated and compared using a feature selection algorithm (SHAP). The performance indicators employed in this study were selected to provide an integrated assessment of how building envelope design parameters influence energy efficiency and indoor environmental quality. Existing literature indicates that airtightness is a significant parameter affecting both energy consumption and indoor thermal performance. Consequently, heating and cooling energy consumption were used to evaluate energy performance, while the PPD (Predicted Percentage of Dissatisfied) and PMV (Predicted Mean Vote) metrics—developed by Fanger—were utilized to assess thermal comfort. Since parameters such as the thermal properties of building envelope components, the window-to-wall ratio, and building orientation can influence not only energy consumption and thermal comfort but also daylight availability and glare conditions, visual comfort performance was incorporated as an objective within the multi-objective optimization problem. Thus, the relevant indicators were evaluated collectively to strike a balance between competing objectives—aiming to reduce energy consumption while ensuring that thermal and visual performance are not compromised.

Figure 1. Methodological flowchart of the study.

Table 1 presents the pseudocode for the proposed multi-stage, multi-objective optimization method. The pseudocode outlines the fundamental algorithmic steps, enabling researchers to understand processes such as modeling, simulation, and programming. It shares common characteristics with standard programming languages and is designed to facilitate easy reading of the algorithm by researchers [27].

Table 1. The specific steps of the surogate model.

1. Create a building model with Honeybee 2. Run simulation:               Calculate Q_heating, Q_cooling, PMV, PPD, sDA, DGP 3. Generate 2000 data points for the parameters using LHS:               X1 ∈ [0.15, 0.60], X2 ∈ [0.80, 2.40], X3 ∈ [0.1, 0.9], X4 ∈ [0.60,5.00], X5 ∈ [0°, 360°] 4. Calculate targets using surrogate equations:               Regression models for Q_heat, Q_cool, PMV, PPD, sDA, DGP 5. Split the data 80/20, normalization:               $\widehat{\mathrm{X}}$ = (X − μ_X)/σ_X; ŷ = (y − μ_y)/σ_y 6. Train SVM for purposes:               f(x) = Σi αi·K(x, xi) + b, K(x, xi) = exp(−γ‖x − xi‖2)         Evaluate:               R2 = 1 − SS_res/SS_tot, RMSE = √(Σ(y − ŷ)2/n) 7. NSGA-II—optimization:               min F(x) = [Q_heat, Q_cool, |PMV|, PPD, −sDA, DGP]T 8. Algorithm:               pop = 100, offspring = 50, SBX(p = 0.9, η = 15), PM(η = 20) 9. Convert Pareto solutions to a matrix:               M ∈ ℝ^(100 × 12) 10. SHAP:               φi(f, x) = Σ_S⊆N\{i} [|S|!(n − |S| − 1)!/n!]·[f(S∪{i}) − f(S)] 11. Permutation importance:               I_j = R2(X, y) − R2(X^(π_j), y) 12. Global importance matrix:               W[i, j] = mean|φi(hedef_j)| 13. Report:               R2 ≥ 0.96 (all objectives), Pareto = 100, the best parameters X1, …, Xn

2.1. Climate of Balikesir and the Building Envelope

The study was conducted in Balıkesir, located in northwestern Türkiye, a region characterized by a mix of Mediterranean and continental climate features. The average annual temperature in Balıkesir is 14.9 °C, with a temperature of 4.8 °C during the coldest month. The highest temperature recorded was 43.2 °C in July. In this region, where winters are cold and summers are hot, the energy demand for heating and cooling is high [28]. Therefore, the performance of the building envelope in the region was evaluated as part of the study. In this study, we integrated the Typical Meteorological Year (TMY)-based EPW file, utilizing Balıkesir’s 2025 climate data, into Ladybug for energy simulations.

We conducted the field study in a north-facing apartment of a 4-story residential building in an urban development area in Balıkesir, surrounded by a few other buildings. The reason for this was to minimize the environmental impact on the apartment. The apartment in question is on the 2nd floor of an apartment building with multiple units per floor, and it has a usable area of 40 m². While preparing the energy model of the apartment, we preserved the geometry and building envelope characteristics. Figure 2 shows the location (a) and the exterior view (b) of the apartment. Table 2 compiles information about the apartment.

Figure 2. The building examined within the scope of the study and its details.

Table 2. Properties of apartment.

Parameter	Unit	Value
Function	-	Residential
Orientation	-	North
Usage Area	m2	40.00
Floor Height	m	3.00
Floor	Number	2
Heating system	-	Combi boiler (Natural Gas)
Cooling System	-	Air conditioning
Wall materail	-	Brick + XPS
Glass	-	Double Glazing

2.2. BlowerDoor Test

The Blower Door Test (BDT) is a method used to detect unwanted air leakage in the building envelope [29]. This method involves sealing the exterior door of the space to be tested with a membrane to prevent airflow and creating a 50 Pa pressure difference between the interior and exterior environments, thereby depressurizing the interior space [5,30,31]. The BDT procedure is based on the ASTM E779 standard [32]. According to this standard, while exterior doors, windows, and vents are closed during the BDT process, all interior doors remain open. Additionally, the dwelling’s heating system and mechanical ventilation units are deactivated during the measurement. The Energy Conservatory (TEC; Oakdale, Minnesota, USA) software records air leakage at 50 Pa pressure in terms of the air change rate per hour (h⁻¹) [33]. Figure 3 shows an image of the BDT being performed on the dwelling.

Figure 3. The BDT process in the residence.

The Blower Door Test (BDT) measurements yield the h⁻¹ value, n₅₀ (air change rate per hour), q₅₀ (air leakage rate per unit of building envelope area), and w₅₀ (air leakage rate per unit of usable floor area). Equations (1)–(3) formulate these values [34].

$$n_{50}={\displaystyle \frac{V_{50}}{V}}[ {\displaystyle \frac{1}{\mathrm{h}}}]$$

(1)

$$q_{50}={\displaystyle \frac{V_{50}}{A_E}}[ {\displaystyle \frac{m^3}{\mathrm{h}.m^2}}]$$

(2)

$$w_{50}={\displaystyle \frac{V_{50}}{A}}[ {\displaystyle \frac{m^3}{\mathrm{h}.m^2}}]$$

(3)

In this study, the n₅₀ value obtained from the BDT serves as a primary input for the model. Python was used to integrate the dwelling’s n₅₀ value into the energy simulations within the parametric model. Consequently, the energy simulation was calibrated using the dwelling’s n₅₀ value, and its reliability was verified against field data. Furthermore, the study defined the n₅₀ value parametrically as the airtightness parameter.

2.3. Building Simulation and Modeling

In this study, simulations for energy consumption, thermal comfort, and daylighting were conducted using a Python interface. The building components of the parametrically modeled residence were defined as shown in Table 1. Python performed the energy simulations for the residence using the EnergyPlus (version 26.2.0) engine. As a result of the EnergyPlus energy simulations, the total monthly heating and cooling loads for the residence in its current state were obtained in kWh. Equations (4) and (5) formulate the monthly heating load and cooling load, respectively.

$$Q_{H,m}=\sum _{t=1}^{N_m}{Q}_{H,t}$$

(4)

$$Q_{C,m}=\sum _{t=1}^{N_m}{Q}_{C,t}$$

(5)

Equation (6) formulates the total energy consumption used in the study.

Q_HVAC,annual = Q_H,annual + Q_C,annual

(6)

The Predicted Mean Vote (PMV) and Predicted Percentage Dissatisfied (PPD) approaches have been used to optimize thermal comfort. Fanger defined the PMV scale as ranging from cold (value of −3) to hot (value of +3). On this scale, ideal thermal comfort conditions are achieved when the PMV value is zero. PMV and PPD values are calculated using Equation (7) and Equation (8), respectively [35,36].

$$\begin{array}{c}\mathrm{PMV}=(0.303{\mathrm{e}}^{-0.036M}+0.028) \left\{\right(\mathrm{M}-\mathrm{W}) –{3.96\mathrm{E}}^{-8}{f}_{\mathit{cl}}[{\left(t_{\mathit{cl}}+273\right)}^4-{\left(\overline{t_r}+273\right)}^4]-f_{\mathit{cl}}{h}_c(t_{\mathit{cl}}-t_a\left)\right]-\\ 3.05[5.733-0.007\left(M-W\right)-P_a]-0.42[\left(M-W\right)-58.15]-0.0173M(5.867-P_a)-0.0014M(34-t_a)\}\end{array}$$

(7)

$$\mathrm{PPD}=100-95e^{({-0.03353\mathit{PMV}}^4-{0.2179\mathit{PMV}}^2)}$$

(8)

where M is Metabolic rate of occupant (W/m² the body surface area); W is mechanical power by occupant, (W/m²); f_cl, (surface area of body with clothes); t_cl, is Clothing surface temperature (°C); t_a is temperature of air (°C); t_r is Mean radiant temperature (°C); P_a is partial pressure of water vapor (Pa).

While balancing the n₅₀ parameter, the study also conducted daylight simulations to ensure that indoor daylight performance was not compromised. Using the parametric model developed, daylight simulations were performed via the software tool DAYSIM (version 3.1). A grid with a spacing of 0.6 × 0.6 m was established within the residential unit, and daylight data were collected at the intersection points of this grid on a work plane located 0.85 m above the floor. In these annual daylight simulations, the study utilized the Spatial Daylight Autonomy (sDA) metric. sDA represents the ratio of the area within a space that receives sufficient daylight throughout the year to the total area of the space. According to the BREEAM standard, this value must be greater than 0.55. In this context, the proportion of the space on the work plane that achieves an illuminance level of 300 lx throughout the year is significant [37]. Equation (9) provides the formula for sDA. Furthermore, to address glare control, the study aimed to minimize the Daylight Glare Probability (DGP). To ensure a high level of glare protection, this metric should be kept below 0.35 [38].

$$\mathit{sD}{A}_{\mathit{Et},r\%}={\displaystyle \frac{A_{\mathit{DAEt}}>r\%}{\mathrm{A}}}\times 100$$

(9)

E_t = target illuminance level, r% = time threshold, A_DAEt > r% = the area that meets the E_t value for at least r% of the year, A = total analysis area.

Model validation: The EnergyPlus model was evaluated against indoor temperature and relative humidity measurements collected using HOBO data loggers (Onset Computer Corporation, Bourne, MA, USA) during two separate 15-day periods: 21 April–5 May and a second period in December. For each period, the measured values were compared with the corresponding simulation results using the coefficient of determination (R²), Mean Bias Error (MBE), and Root Mean Square Error (RMSE) [39]. The use of measurements from different periods of the year for building model evaluation has also been reported in previous studies. Roberti et al. [40] calibrated an EnergyPlus model using summer measurements and validated it using winter measurements, while Palaić et al. [41] used separate winter and summer periods for calibration and validation.

$$R^2=1-{\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}{\left(y_i-{\widehat{y}}_i\right)}^2}{{\displaystyle {\sum }_{i=1}^n}{\left(y_i-\overline{y}\right)}^2}}$$

(10)

$$\mathit{MBE}(\%)={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^{\mathit{Np}}}(\mathit{mi}-\mathit{ms})}{{\displaystyle {\sum }_{i=1}^{\mathit{Np}}}\left(\mathit{mi}\right)}}$$

(11)

$$\mathit{CV} \mathit{RMSE}(\%)={\displaystyle \frac{\surd {\displaystyle {\sum }_{i=1}^{\mathit{Np}}}{\left(\mathit{mi}-\mathit{ms}\right)}^2/\mathit{Np}}{\mathit{mp’}}}$$

(12)

2.4. Machine Learning-Based Optimization

In the optimization process, an ML-based surrogate model was developed to reduce computational costs. Using a parametric model, 2000 distinct simulations—covering energy, thermal comfort, and daylight—were performed for selected design parameters via Python. Data generated during these performance simulations were recorded. Recently, researchers have employed hybrid models that combine surrogate models with optimization algorithms to address multi-objective optimization problems [42,43,44]. The application of surrogate models and machine learning techniques contributes significantly to the field of building performance optimization [44,45,46]. Machine learning algorithms—specifically Support Vector Machine (SVM), Artificial Neural Network (ANN), and XGBoost—were utilized to predict the outcomes based on the collected data. Subsequently, the performance of these algorithms was compared using performance metrics, cross-validation, robustness testing, and independent test datasets.

The baseline parameters for the machine learning algorithms, as well as those optimized using Optuna, are presented in Table 3. The dataset was divided into a training set and a test set, with an 80/20 split [47,48]. The datasets used in the training and testing processes of the ANN, SVR, and XGBoost algorithms are not derived from actual measured data from a real building; rather, they are obtained using the EnergyPlus simulation tool. The results concerning prediction performance do not represent the actual performance of the building itself; instead, they demonstrate the alignment of the machine learning-based models with the simulation outputs. Within the Python programming environment, the Scikit-learn library (version 1.9.1) was used for Artificial Neural Network (ANN) and Support Vector Regression (SVR) models, while the XGBoost library (version 3.4.1) was used for the Extreme Gradient Boosting (XGBoost) model. These models were implemented in the Google Colab (Python 3.12) environment on a system equipped with an Intel Core i7-8750 processor and 8 GB of RAM.

Table 3. Model parameters used in this study.

Algorithm	Parameters	Optimized Hyperparameters
SVR	Model type (Support Vector Regression), kernel function (Radial Basis Function, RBF), regularization parameter C (10.0), epsilon-insensitive loss parameter (0.05), kernel coefficient (gamma = scale).	Kernel: RBF; regularization parameter C: 0.1–1000; gamma: 0.0001–1; epsilon: 0.001–0.3.
ANN	Model type (Multilayer Perceptron), number of hidden layers (2), neurons per hidden layer (128, 64), activation function (ReLU), optimizer (Adam), initial learning rate (0.001), L2 regularization parameter (0.001), maximum iterations (2000), early stopping (True), validation fraction (0.15), number of iterations without improvement (30), random seed (42).	Hidden-layer architecture: (64, 32), (128, 64), or (128, 64, 32); L2 regularization (alpha): 10−6–10−2; learning rate: 10−4–10−2; activation: ReLU; optimizer: Adam; maximum iterations: 400; early stopping: True; validation fraction: 0.15; patience: 15.
XGBoost	Number of estimators (500), maximum tree depth (5), learning rate (0.05), subsample ratio (0.9), column sampling ratio (0.9), objective function (squared error regression), tree method (histogram-based), parallel processing (n_jobs = −1), random seed (42).	Number of estimators: 100–300 (step 100); maximum depth: 2–8; learning rate: 0.01–0.20; subsample: 0.70–1.00; column sampling: 0.70–1.00; objective: squared error regression; tree method: histogram-based.

Algorithm evaluations were conducted using Optuna for hyperparameter optimization, 5-fold cross-validation (CV) for model validation, 5-repeated 5-fold CV for robustness analysis, and R², RMSE, and MAE as performance metrics. This approach ensured the selection of the most effective model for the non-linear regression relationship between the objective function and the design variables [49,50,51]. This regression relationship serves as the fitness function within the NSGA-II heuristic algorithm, which is employed to solve the multi-objective optimization problem. NSGA-II, an evolutionary-based algorithm, establishes a balance among the parameters selected within the scope of the study. Through its Pareto-front solutions, NSGA-II facilitates the selection process within the design space [52,53]. In this context, an SVM-based surrogate model—demonstrating the highest performance—was utilized as the objective function for NSGA-II, thereby reducing computational time.

The Identifying design variables: This study considers the interaction with the indoor visual environment while balancing architectural design parameters that affect energy consumption and the airtightness of the building envelope. The parameters related to the building envelope were selected to simultaneously evaluate thermal comfort, energy performance, and daylighting. The effective use of daylight and the control of glare within the interior space formed the basis for assessing indoor visual conditions. Consequently, indoor visual conditions were not overlooked while balancing energy performance. The architectural design parameters (Table 4) addressed in the study were selected in alignment with the established objectives.

Table 4. Design variables and their ranges.

Parameter	Unit	Boundary
X1: Wall U-value	W/m2K	[0.15–0.60]
X2: Glazing U-value	W/m2K	[0.80–2.40]
X3: WWR	%	[0.30–0.80]
X4: Airtightness (n50)	h−1	[0.50–5.00]
X5: Orientation	°	[0.00–360]

In a multi-objective optimization process, an objective function must be established for each optimization goal. The Support Vector Machine (SVM) algorithm is employed to estimate each of these goals. A non-linear model, developed to represent the relationship between the design variables of the building facade and each optimization objective, serves as the objective function. This objective function can be expressed through the output function of the SVM algorithm. Since a reduction in indoor air leakage is directly proportional to a decrease in the Window-to-Wall Ratio (WWR), the study incorporates the performance metrics PMV, PPD, sDA, and DGP—ensuring that thermal and visual comfort are not compromised while balancing operational energy consumption. As the optimization problem is framed as a minimization task, 1/sDA is used as the objective function for sDA, and |PMV| is used as the objective function for PMV. The multi-objective optimization problem is defined as shown in Equation (13):

$$\left\{\begin{array}{c}\min f_1\left(\mathrm{SVM}\left(x_1,x_2,x_3\dots ,x_5\right)\right)\\ \min f_2\left(\mathrm{SVM}\left(x_1,x_2,x_3\dots ,x_5\right)\right)\\ \min f_3\left(\mathrm{SVM}\left(x_1,x_2,x_3\dots ,x_5\right)\right)\\ \min f_4\left(\mathrm{SVM}\left(x_1,x_2,x_3\dots ,x_5\right)\right)\\ \min f_5\left(\mathrm{SVM}\left(x_1,x_2,x_3\dots ,x_5\right)\right)\\ \min f_6\left(\mathrm{SVM}\left(x_1,x_2,x_3\dots ,x_5\right)\right)\\ X1\in \left[0.15, 0.60\right], X2\in \left[0.8, 2.4\right], X3\in \left[0.10, 0.90\right], \\ X4\in \left[0.50, 5.00\right], X5\in \left[0, 360\right]\end{array}\right.$$

(13)

2.5. Identification of Influential Parameters

The SHAP (SHapley Additive ExPlanations) algorithm was used to identify and visualize influential parameters. SHAP employs a game-theoretic approach to interpret machine learning models [54,55]. Machine learning models often function as “black boxes”—taking certain features as input and producing predictions as output. Model interpretability plays a significant role in the field of artificial intelligence [56]. When researchers gain greater insight into how a machine learning model arrives at its final output, they are better equipped to interpret the problems at hand compared to using algorithms that act as opaque, “closed” boxes [57] (Figure 4).

Figure 4. Approaches of explainable models and black-box models to problems.

Explainable AI tools can be used to provide clear and understandable interpretations regarding the model’s output. The Shapley value is calculated as shown in Equation (14):

$${\varnothing }_i= {\sum }_{S\subseteq F\setminus \left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(\left|F\right|-\left|S\right|-1\right)!}{\left|F\right|!}}\left[F_{S\cup \left\{i\right\}}\left(x_{S\cup \left\{i\right\}}\right)-f_S\left(x_S\right)\right]$$

(14)

Here, F represents the set of all features, S represents the subset of features excluding the i-th feature, and f(S ∪ {x_i}) − f(S) represents the difference between predictions that include and exclude the i-th feature.

3. Results and Discussion

3.1. Blower Door Test Results

In this study, we conducted Blower Door Tests (BDTs) at four different times—09:00, 12:00, 15:00, and 18:00—on 20 March 2026, in a north-facing apartment within a multi-unit building in Balıkesir, taking into account fluctuations in wind speed and air temperature throughout the day. As a result of the BDT, we determined the air change rate (n₅₀) at a pressure of 50 Pa. Table 5 presents the n₅₀ values obtained at these four different times. The values in the table indicate that the dwelling maintains a consistent airtightness profile based on the measurements taken at various times of the day.

Table 5. BDT results for the dwelling.

Time	09.00	12.00	15.00	18.00
n50	1.66	1.69	1.58	1.56

In the evaluation process, we used the average of the n₅₀ values (1.6225) obtained during the BDT procedure. Table 6 presents airtightness limit values for residential buildings in various countries. When considering the limit values in the table, the dwelling under investigation meets the requirements of the TS 825 [58], Bulgaria Passive House, Austria Passive House, and Czech Republic TNI 730330 standards [59,60]. Furthermore, the results obtained with low standard deviation throughout the BDT process enhance the reliability of the analysis performed.

Table 6. Airtightness limit values in different countries [58,59,60].

Standard	Country	n50
TS 825	Türkiye	n50 < 2 (many apartments per floor)
		n50 < 4 (one apartment per floor)
Passive House	Europe	n50 < 0.6
	Bulgaria	n50 < 2
	Austria	n50 < 3
TNI 730330	Czech Republic	n50 < 4.5
	Denmark	n50 < 1.5
	Germany	n50 < 1.5

3.2. Simulation Results

The results of code-based energy, thermal comfort, and daylight simulations conducted via Python are presented in Figure 5 and Table 7.

Figure 5. Building performance simulation results for the existing condition.

Table 7. Simulation results.

	Month	Heating (kWh)	Cooling (kWh)	Total (kWh)	PMV	PPD (%)	sDA (%)	DGP
0	Jan	820.4	0.0	820.4	−0.85	20.4	38.2	0.22
1	Feb	710.1	0.0	710.1	−0.72	16.1	42.1	0.24
2	Mar	530.2	0.0	530.2	−0.45	9.3	46.5	0.26
3	Apr	240.5	15.2	255.7	−0.15	5.5	52.0	0.29
4	May	45.0	115.4	160.4	0.12	5.2	56.4	0.31
5	Jun	0.0	310.8	310.8	0.48	9.8	58.2	0.33
6	Jul	0.0	450.2	450.2	0.65	14.1	59.0	0.34
7	Aug	0.0	425.6	425.6	0.61	12.8	57.5	0.32
8	Sep	12.3	210.1	222.4	0.25	6.1	53.1	0.29
9	Oct	210.4	45.3	255.7	−0.10	5.1	47.0	0.27
10	Nov	512.2	0.0	512.2	−0.55	11.2	41.0	0.24
11	Dec	769.1	0.0	769.1	−0.78	17.8	36.5	0.21

Energy performance: Table 7 displays the monthly heating and cooling energy results obtained from simulations performed using the EnergyPlus engine. An examination of the table reveals that the energy demand for heating peaks in January (820.4 kWh), while the energy demand for cooling peaks in July (450.2 kWh). Based on the annual total energy requirements, the total heating energy is 3850.2 kWh, and the total cooling energy is 1572.6 kWh. The fact that heating energy accounts for approximately 71% of the total annual energy requirement indicates that the heating load is the dominant factor for the residence under study. This dominance of heating-related energy consumption is a consequence of the residence’s north-facing orientation.

Thermal comfort performance: As shown in Figure 5 and Table 7, PMV values vary between −0.85 and +0.65 over a 12-month period. Correspondingly, PPD values range from 5.1% to 20.4%. The simulation results indicate that thermal discomfort in the residence occurs primarily during the summer and winter months. Specifically, the results highlight thermal dissatisfaction due to cold weather during the winter, and dissatisfaction due to hot weather during the summer. PMV values show an upward trend during the summer months, reaching their peak in August. An examination of the monthly average PMV values in the chart reveals that, during the winter months (December, January, and February), the PMV values range approximately from −0.72 to −0.85, falling outside the ±0.5 comfort range commonly used under the ASHRAE 55 standard [61]. In the summer months (June, July, and August), these values range approximately from +0.48 to +0.65; notably, the results for July and August fall outside the comfort range. Furthermore, it has been observed that thermal conditions are more favorable during the transitional seasons compared to the summer and winter months. Given the levels of discomfort observed during the summer and winter periods, it is essential that design variables related to the building envelope account for thermal comfort in addition to energy performance.

Visual Comfort Performance: Simulation results indicate that both the level of daylight utilization and the probability of glare undergo seasonal variations over a 12-month period. The sDA (spatial Daylight Autonomy) values were found to range from a minimum of 36.5% to a maximum of 59.0%, while DGP (Daylight Glare Probability) values varied between 0.21 and 0.34. The sDA values are relatively lower during the winter season compared to other periods, suggesting that natural lighting is insufficient during this time. In this same period, DGP values range from 0.21 to 0.24, indicating a lower probability of glare than in other seasons. Conversely, the situation during the summer months presents the opposite pattern. Daylight analyses conducted using Python yielded an sDA value of 55.5% for the north-facing residential unit. This meets the 55% threshold for sDA as per the BREEAM standard [37]. Overall, the simulation results show that the performance for all three objectives varies according to seasonal conditions. This necessitates an evaluation that accounts for the interdependencies among the performance indicators selected for multi-objective optimization.

Model Validation: This process involves validating the simulation model for indoor air temperature by comparing it against actual data. The study demonstrates the validation of the simulation model for indoor air temperatures during April and May (Figure 6) and December (Figure 7). The performance simulation model developed in this study has been proven reliable according to ASHRAE Guideline 14, yielding an R² value of 98.40%, an MBE of 0.87%, and a CV(RMSE) of 2.72% for the period covering April and May, and an R² value of 99.25%, an MBE of 0.63%, and a CV(RMSE) of 1.63% for the period covering December.

Figure 6. Validation of actual and simulated indoor temperature values for the months of April and May.

Figure 7. Validation of actual and simulated indoor temperature values for the month of December.

3.3. AI-Based Optimization Results

To develop an ML-based surrogate model, 2000 performance simulations were conducted using various combinations of design variables, and both the design variables and the resulting performance metrics from these simulations were recorded. The machine learning algorithms employed in this study reduced the computational cost of the optimization process by providing predictions. The predictive accuracy of the surrogate model directly influences the reliability of the results obtained from the optimization [62]. Table 8 presents a comparison of the performance metrics for three different algorithms. An examination of the predictive performance of these algorithms across the various objectives reveals that all three achieve high R² values (0.96–0.99). The fact that the lowest R² values among the machine learning algorithms are associated with the sDA objective highlights the inherent difficulty in predicting this specific target. In this study, the machine learning-based surrogate models used as objective functions for optimization were not employed to predict performance data measured from an actual building. Instead of measured data from a real building, data obtained from EnergyPlus simulations were used as training and testing sets for the machine learning phase. Consequently, the statistical performance indicators used to evaluate the predictive capabilities of the ANN, SVR, and XGBoost algorithms reflect the success of these surrogate models in replicating the outputs generated by the EnergyPlus simulation engine. In multi-objective optimization processes, the continuous execution of simulation tools is typically time-consuming, and optimization algorithms often require the simultaneous evaluation of numerous criteria [46,63]. Therefore, machine learning-based surrogate models are utilized as objective functions within multi-objective optimization processes, as they provide high-accuracy predictive models that help overcome the drawbacks associated with high computational costs [25,64,65].

Table 8. Comparison of performance metrics for three different algorithms.

Algorithm	Objective	R2	RMSE	MAE
SVR	Heating_Load_kWh	0.9986	2.5857	1.9091
	Cooling_Load_kWh	0.9982	3.0073	2.0328
	PMV	0.9984	0.0050	0.0036
	PPD	0.9986	0.0782	0.0600
	sDA_Percent	0.9650	3.3716	2.6332
	DGP	0.9942	0.0036	0.0027
ANN	Heating_Load_kWh	0.9979	3.1960	2.5061
	Cooling_Load_kWh	0.9986	2.6533	2.0178
	PMV	0.9988	0.0044	0.0035
	PPD	0.9986	0.0783	0.0624
	sDA_Percent	0.9700	3.1242	2.5050
	DGP	0.9993	0.0013	0.0010
XGBoost	Heating_Load_kWh	0.9936	5.6288	4.3222
	Cooling_Load_kWh	0.9958	4.6295	3.2779
	PMV	0.9961	0.0078	0.0060
	PPD	0.9957	0.1357	0.0963
	sDA_Percent	0.9700	3.1217	2.4715
	DGP	0.9986	0.0018	0.0012

Figure 8, Figure 9 and Figure 10 present the predictive performance of the surrogate model. The figures show that the difference between the values predicted by machine learning models and the simulation values is very small. Consequently, the regression equation, by aligning with the actual conditions, effectively captures the relationship between the design parameters and the six distinct objectives [63,66]. An examination of the error metrics reveals that no single algorithm consistently outperformed the others across all objectives (Table 8). SVR yielded the lowest error values for heating load and PPD; ANN performed best for cooling load, DGP, and PMV; and XGBoost achieved the lowest error for sDA. Given that relying solely on these performance metrics is insufficient to determine the most effective algorithm, the study incorporates further analyses, including cross-validation, robustness testing, and evaluations using independent test datasets. These analyses encompass the optimization of 18 separate models—covering 3 different algorithms and 6 distinct objectives—with each model undergoing training and robustness assessments based on 20 Optuna trials and 5-fold cross-validation.

Figure 8. Prediction performance of the SVM for different objectives.

Figure 9. Prediction performance of the ANN for different objectives.

Figure 10. Prediction performance of the XGBoost for different objectives.

Cross-Validation (CV) is a technique that helps in estimating the robustness and performance of models. Primarily, it is used to predict how a model will perform in a production environment; thus, it serves as a crucial step in accounting for the potential variability of data in real-world scenarios [67]. Table 9 presents the results of algorithm selection based on repeated cross-validation. The SVR algorithm demonstrated the best performance for cooling load, heating load, PMV, and PPD objectives, while the XGBoost algorithm performed best for DGP and sDA objectives. Similarly, in the existing literature on building performance optimization, SVM-based algorithms have also shown high performance [25,68]. In this study, the ANN algorithm did not rank first for any of the objectives. Consequently, the SVR algorithm exhibited superior performance compared to the other algorithms. The final independent test results for this study are provided in Appendix A. The fact that the 5-fold CV, repeated cross-validation, and independent test results are closely aligned indicates that the models maintain consistent performance across different data splits. Furthermore, the individual validation and error plots for all algorithms also confirm these results (Appendix B).

Table 9. The best algorithm selected via cross-validation on the training data.

Objective	Algorithm	Repeated CV R2 Mean	Repeated CV RMSE Mean	Test R2	Test RMSE
Cooling_Load_kWh	SVR	0.9998	0.9706	0.9998	0.9002
DGP	XGBoost	0.9999	0.0004	0.9999	0.0004
Heating_Load_kWh	SVR	0.9999	0.8078	0.9999	0.6174
PMV	SVR	0.9994	0.0029	0.9995	0.0028
PPD	SVR	0.9993	0.0551	0.9994	0.0530
sDA_Percent	XGBoost	0.9735	2.9332	0.9738	2.9191

In buildings, there is a trade-off between energy performance and daylight performance. For this reason, this study adopts a Pareto-front-based multi-objective optimization (MOO) process that minimizes building energy consumption while maximizing indoor daylight and thermal performance. Figure 11 presents the 6-dimensional Pareto-front obtained from the optimization. An examination of the Pareto-front reveals that the selected alternatives achieve optimal performance across multiple indicators simultaneously, including daylight, visual comfort, thermal comfort, and energy performance. Among the proposed models, the best-performing option requires 2760.82 kWh for annual heating and 1389.33 kWh for cooling, with a PMV value of 0.431, a PPD of 8.88%, an sDA of 65%, and a DGP of 0.335.

Figure 11. Optimum result obtained from multi-objective optimization.

Table 10 presents 10 different proposed models. These models demonstrate that lower thermal transmittance values for walls and windows lead to a reduction in total energy consumption. An examination of the window-to-wall ratio (WWR) reveals various alternatives ranging from 0.3 to 0.8. The study identifies n₅₀ values within the range of 0.60–5.00 h⁻¹ offering a variety of proposals with both favorable and less favorable n₅₀ values. Regarding building orientation, it was found that proposals oriented primarily toward the south or southwest yield optimal results. Furthermore, significant differences are observed in terms of energy consumption, thermal comfort, and visual comfort performance.

Table 10. Design solutions for the proposed models.

ID	X1	X2	X3	X4	X5	Heating (kWh)	Cooling (kWh)	|PMV|	PPD (%)	sDA (%)	DGP
1	0.506	2.385	0.8	3.84	225	2872.47	1373.82	0.431	8.88	65.00	0.335
2	0.596	2.090	0.8	4.80	360	2933.57	1345.23	0.301	6.92	65.74	0.344
3	0.153	2.310	0.6	0.62	180	2662.48	1308.00	0.709	15.77	49.34	0.299
4	0.499	2.099	0.8	0.88	315	2740.53	1340.18	0.529	10.82	71.94	0.349
5	0.218	0.800	0.3	4.04	270	2727.37	1138.38	0.226	6.07	30.61	0.257
6	0.150	1.183	0.3	0.60	90	2581.48	1165.02	0.485	10.12	20.04	0.228
7	0.600	0.800	0.7	5.00	90	2834.65	1203.39	0.040	5.20	47.12	0.305
8	0.213	1.883	0.8	5.00	270	2860.51	1307.15	0.396	8.30	71.74	0.355
9	0.150	1.207	0.3	2.12	90	2640.65	1161.34	0.422	8.74	15.67	0.231
10	0.555	0.865	0.7	4.59	90	2822.37	1203.68	0.069	5.07	47.65	0.304

While some solutions offer advantages in terms of energy consumption (such as solution 6), others—through different combinations—yield better results regarding thermal or visual comfort. For instance, an examination of solution 1 reveals a design with a high transparency ratio, characterized by an 80% window-to-wall ratio and a glazing U-value of 2.385 W/m²K. Although the sDA (spatial Daylight Autonomy) value of 65% is favorable for daylighting, the total energy consumption of 4246.29 kWh represents a poor outcome in terms of energy performance. Meanwhile, the values of |PMV| = 0.431 and PPD = 8.88% indicate that the level of thermal comfort falls within an acceptable range. In terms of energy performance, solution 6 is the most favorable, whereas solution 2 exhibits the highest energy consumption; there is a difference of approximately 12.44% in energy consumption between these two options. Solution 7 provides the best thermal comfort, while solution 3 performs the worst in this regard. These findings highlight that low energy consumption alone does not guarantee optimal thermal comfort conditions. Regarding visual comfort, solutions 4 and 8 achieve the best results, boasting the highest sDA values. Ultimately, evaluating different objectives—such as energy consumption, thermal comfort, and visual comfort—independently during the building design process is not an appropriate approach. A design alternative that performs well for one objective may yield undesirable results for another, which underscores the importance of multi-objective optimization.

3.4. Sensitivity Analysis Results

This study investigated the individual impact of each parameter on performance indicators using standardized regression coefficients and visualized these findings through partial dependence plots. The SHAP method was employed for feature selection and to ensure explainability; SHAP serves as a robust tool for elucidating the contributions of specific features to a model. Figure 12 illustrates the influence of various parameters on heating energy. The figure reveals that parameter X5 is the most significant factor affecting the heating load, exhibiting an inverse relationship with it. This is followed by parameter X4, which shows a positive correlation with the heating load. The results indicate that the airtightness of the building envelope is among the most influential parameters affecting the indoor heating load.

Figure 12. Effect of parameters on heating load.

Figure 13 presents the effect of the parameters on cooling energy. According to the figure, it was determined that X5 is the most influential parameter regarding cooling load and that it is inversely proportional to the cooling load. This parameter is followed by parameters X2 and X3, both of which exhibit a negative relationship with the cooling load.

Figure 13. Effect of parameters on cooling load.

Figure 14 presents the effect of the parameters on PMV. According to the figure, it was determined that X5 is the most influential parameter regarding PMV and that it has an inverse relationship with PMV. The building envelope airtightness value was the second-most influential parameter affecting PMV; as the value of this parameter increases, PMV is negatively affected.

Figure 14. Effect of parameters on PMV.

Figure 15 presents the effect of parameters on PPD. According to the figure, it was determined that the most effective parameter on PPD was X5 and that this parameter was directly proportional to PPD. Building envelope airtightness value was the 2nd most effective parameter on PPD. As the value of this parameter increases, PPD is negatively affected.

Figure 15. Effect of parameters on PPD.

Figure 16 presents the effect of parameters on sDA. According to the figure, orientation is the most influential parameter regarding daylight access to the space. This is followed by WWR. It was determined that the airtightness of the building envelope has a minimal effect on indoor daylight performance.

Figure 16. Effect of parameters on sDA.

Figure 17 presents the effect of parameters on DGP. According to the figure, orientation is the most significant parameter affecting the probability of glare within the space. The second-most effective parameter for indoor glare control is the WWR.

Figure 17. Effect of parameters on the DGP.

4. Conclusions and Recommendations

This study proposes a methodology that integrates air-tightness measurements—conducted in an actual residential dwelling—with building performance simulations, machine learning, and multi-objective optimization techniques. The proposed approach aims to generate balanced design solutions by simultaneously evaluating the impacts of building envelope design parameters on energy consumption, thermal comfort, and visual comfort. To achieve this, surrogate models—based on three different machine learning models (including Support Vector Machines and Artificial Neural Networks) and a metaheuristic algorithm (Non-dominated Sorting Genetic Algorithm II, or NSGA-II)—were employed to identify optimal combinations of building design variables. The selection of the most effective machine learning model was supported by analyses such as independent testing, validation, and the use of performance metrics frequently cited in the literature. The Support Vector Machine algorithm, which demonstrated effective and reliable predictive performance, was identified as the most suitable model and was subsequently used to generate the optimal solutions.

The multi-objective optimization, performed using NSGA-II, resulted in a 21.69% reduction in energy consumption, a 17.2% improvement in visual comfort, and a 20% improvement in thermal comfort compared to the baseline (existing) conditions. These findings underscore the importance of simultaneously considering energy efficiency and user comfort. Analyses conducted using the Shapley Additive Explanations (SHAP) method revealed that the relative influence of design parameters varies across different performance indicators. Consequently, this study highlights the significance of accounting for the distinct impacts of various parameters when optimizing for energy consumption, thermal comfort, and visual comfort within a multi-objective optimization framework. While parameters such as the Window-to-Wall Ratio (WWR) and building orientation are significant for visual comfort, variables related to infiltration and the thermal transmittance of building components are key to enhancing energy efficiency and thermal comfort. These results demonstrate that the same design parameters do not exert an equal influence on different performance objectives. In this context, a multi-objective optimization process provides a framework for evaluating the impacts of performance improvements across various goals, rather than focusing on a single performance indicator.

Based on these findings, shaping a design by simultaneously considering performance indicators that exist in a trade-off relationship enables more informed decision-making. Prioritizing one comfort criterion while neglecting others fails to achieve optimal indoor conditions. This study distinguishes itself from existing research by incorporating verified infiltration data and examining the relationship between thermal comfort and energy efficiency, while also accounting for daylighting and glare control. Alongside its original contributions, the study also presents certain limitations. The research focuses on a single residential building, which restricts the generalizability of the findings to other building typologies. Furthermore, the machine learning models employed in this study were trained and tested using output data generated by the EnergyPlus simulation engine, which constitutes another limitation. The high predictive accuracy of these machine learning models reflects a strong alignment with simulation outputs rather than with actual, real-world building performance. Consequently, the performance improvements achieved through multi-objective optimization—regarding energy consumption, thermal comfort, and visual comfort—should be interpreted as estimates derived from simulation-based modeling. The use of three distinct machine learning models alongside the NSGA-II algorithm in multi-objective optimization processes limits the ability to compare the performance of various modeling and optimization approaches. While this study confirms the reliability of the SVM and NSGA-II optimization algorithm in optimizing energy, thermal, and visual comfort performance, there remains a need for design optimization that addresses additional objectives, such as life-cycle costs. In this study, equal weighting coefficients were assigned to the objective functions within the optimization process; however, in multi-objective optimization problems, assigning different weights to specific objective functions could yield solutions tailored to user preferences. Furthermore, extending the scope to examine multiple buildings—rather than focusing solely on a single new residential building—would be a compelling direction for future research. Future studies could also develop models that produce more realistic results by creating a “digital twin” of a space, utilizing IoT-based sensors to collect real-time data on indoor pressure, temperature, and glare.