Interpretable machine learning framework for climate-resilient architectural design optimization

Xinyi Zhang

doi:10.1080/17452007.2026.2702035

Interpretable machine learning framework for climate-resilient architectural design optimization

2026-07-15

0

PRE

AI

DOI：10.1080/17452007.2026.2702035

Original

Original request for help

Save

Abstract

En 中文

Enhancing climate resilience in the built environment requires data-driven approaches to optimize passive climate-responsive design strategies, and traditional Chinese dwellings offer valuable references for adaptive architectural performance across climates. This study develops an interpretable machine learning framework for performance-driven and climate-resilient architectural design optimization by integrating building performance simulation with Extreme Gradient Boosting (XGBoost), Bayesian optimization, and SHapley Additive exPlanations (SHAP). While conventional building performance simulation tools (e.g. DesignBuilder and EnergyPlus) support parametric analysis and remain essential for detailed evaluation, the interpretation of complex, nonlinear interactions among multiple design parameters is often implicit and difficult to systematically extract. Our framework complements these methods by (1) providing a targeted surrogate model for rapid prediction of PMV and UDI within the Pareto-optimal high-performance design region; and (2) employing SHAP to quantitatively uncover and explain the nonlinear influences and synergistic effects of key design parameters, such as the combined impact of patio size and orientation, delivering interpretable and design-oriented insights that are difficult to derive directly from traditional simulation alone. A case study of a traditional residence in Guilin, China, demonstrates applicability to early-stage design decision-making. The optimized XGBoost model achieves a cross-validated R² of 0.9968, and TPE-based Bayesian optimization outperforms conventional grid search and Gaussian process approaches in efficiency. SHAP analysis identifies patio size, orientation, and buffer space as dominant drivers of daylight and thermal performance. The study contributes interpretable, data-driven insights that extend conventional building performance modeling and support climate-resilient design.

Keywords:

Architectural design optimization

climate-resilient design

passive strategies

traditional Chinese dwellings

interpretable machine learning

Interpretable machine learning framework for climate-resilient architectural design optimization

Abstract

Journal

Organization

Cited Papers