Investigation of the relationships between enclosed architectural spatial forms and indoor–outdoor thermal comfort based on explainable machine learning and optimization algorithms

To address the climate change-induced challenges to thermal comfort, in this study, an innovative framework that integrates multiobjective optimization (MOO) and explainable machine learning (ML) is introduced to investigate the influencing mechanisms of spatial morphology on indoor–outdoor thermal comfort (IOTC) and to provide quantitative guidance for high-performance design. An MOO model was developed using a genetic algorithm (GA), with nine morphological parameters—six related to building form and three related to courtyard form—serving as decision variables to simultaneously optimize predicted mean vote (PMV) and universal thermal climate index (UTCI) values in summer (minimization) and winter (maximization). An ensemble ML model, enhanced through Bayesian hyperparameter optimization, was employed for thermal comfort prediction, among which XGBoost demonstrated superior performance. Furthermore, SHapley Additive exPlanations (SHAP) analysis was applied to quantitatively interpret the contributions and interaction effects of key morphological parameters. The results show that three parameters—the building shape index (BSI), morphological coefficient (FSC), and building-to-courtyard ratio (BCR)—collectively account for over 76% of the thermal comfort variation, with a particularly notable interaction effect between the BSI and BCR. By transcending the limitations of conventional black-box optimization, the proposed framework can be used to establish a decision-support tool that in which global optimization capability and mechanistic interpretability are integrated. The results not only deliver optimal design solutions but also elucidate the underlying reasoning, thereby offering a data-driven, interpretable, and scalable methodology for climate-adaptive architectural design.