Physics-informed graph attention networks for multi-objective optimization of double skin façades: a climate-responsive framework for operational carbon reduction

This study introduces a physics-informed and climate-responsive optimization framework of Double Skin Façade (DSF) design to reduce the operational CO 2 emissions and preserve thermal and ventilation functions. The method combines building energy simulations and a Physics-Informed Graph Attention Network (PI-GAT) surrogate model and a multi-objective NSGA-II algorithm. The PI-GAT model outperforms conventional machine learning methods (ANN, RF, and XGBoost) by achieving higher accuracy and improved robustness, particularly under nonlinear and extreme design conditions, due to the incorporation of physical constraints governing airflow and heat transfer. The model is tested on two opposite climates: Tashkent (cold-dominant) and Najran (hot-arid). Findings show that optimum DSF schemes are capable of decreasing operational CO 2 emissions by 13–23% and 8–20% in comparison with baseline scenarios, respectively. Pareto analysis shows that there are unique climate-dependent trade-offs in which ventilation needs to be tightly regulated in cold climates but improved in hot climates to attain carbon-efficient operation. Overall, the proposed PI-GAT-NSGA-II framework is a credible and computationally efficient method of climate-adaptive and low-carbon design of façades.