Explainable machine learning for passive design: Early-stage building energy reduction in hot-arid climates

Accurately predicting heating and cooling loads is crucial for sustainable building design, particularly in early stages when passive strategies can reduce reliance on active systems. This study develops an interpretable machine learning (ML) framework to assess the impact of geometric variables (floors, orientation, area, aspect ratio) and envelope parameters (SHGC, WWR, U-value, albedo, VT) on energy demand in hot-arid climates. A simulation dataset was generated using EnergyPlus. Six ML models Linear Regression, Support Vector Machine, K-Nearest Neighbors, Decision Tree, Random Forest, and Gradient Boosting were trained and compared. Gradient Boosting achieved the best accuracy at both scales. For geometry, it reached R² = 0.9909 with very low errors (MAE = 0.68, RMSE = 0.90 kWh/m²·year), identifying Number of Floors and Orientation as most influential. For envelope parameters, accuracy was lower (R² = 0.8290, MAE = 10.97, RMSE = 8.76), with SHGC dominating by increasing cooling loads. Linear models underestimated these effects, while ensemble methods captured nonlinear interactions. These findings emphasize the role of passive design in lowering HVAC demand and environmental impacts. The dual-scale framework functions as a sensitivity analysis tool, helping decision-makers prioritize variables and supporting optimization workflows for sustainable design.