Climate-Generalizable Energy Prediction in PCM-Integrated Building Envelopes: A Physics-Informed Machine Learning Framework for Sustainable Envelope Design

Phase change materials (PCMs) offer potential for passive thermal regulation in building envelopes through latent heat storage; however, their effectiveness remains strongly climate-dependent, and configurations optimized for one region often underperform in others. Existing evaluation approaches rely largely on location-specific simulations or surrogate models with limited climate transferability. This study develops a physics-informed, climate-aware machine-learning framework to assess PCM-integrated wall assemblies across diverse climates. A structured dataset of 720 EnergyPlus simulations was generated across nine PCM materials, ten ASHRAE climate zones, two placement configurations, and four thickness levels using automated model generation and batch simulation through Eppy-based workflows. Ensemble-based models (XGBoost, LightGBM, CatBoost, Random Forest) were trained under climate-grouped validation to predict total annual energy consumption, peak cooling demand, and peak heating demand. The models achieved high predictive accuracy for total annual energy use (R2 ≈ 0.98–0.99) and peak cooling demand (R2 ≈ 0.93–0.96), outperforming statistical, climate-only, and PCM-agnostic baselines. In contrast, peak heating demand showed low predictability (R2 ≤ 0.26), indicating limited sensitivity to PCM parameters under the studied configuration. These results demonstrate that climate-aware validation enables defensible cross-climate PCM assessment, supporting energy demand reduction and sustainable envelope design decisions aligned with global building decarbonization goals.