Deep learning-based surrogate modeling for optimal control of climate-responsive adaptive faCade systems

The adaptive façade (AF), an active system for precise environmental control, has demonstrated substantial potential in enhancing energy efficiency and thermal comfort. With increasing demands for adaptive open-space configurations, modular AF systems featuring individually controllable units have become particularly important. However, conventional numerical simulation methods are constrained by computational cost and cannot meet the requirements of real-time AF control strategies.

To address this, we propose a rapid prediction framework integrating a GAN-based agent model (pix2pixHD) with the NSGA-II algorithm. By developing a parametric AF system, the framework achieves real-time prediction of daylight autonomy (DA) and thermal autonomy (TA), facilitating multi-objective optimization to generate instant feedback for AF opening and closing patterns. Model validation involves accuracy analysis and scenario assessment to evaluate the predicted climatic responses on indoor thermal and lighting conditions.

Compared to traditional surrogate models based on structured numerical predictions, the proposed GAN-based framework provides superior visual interactivity and interpretability, significantly improving efficiency in architectural design evaluation and decision-making processes.