Multi-Model AI Solutions for Window Configuration in Complex Architectural Scenarios with Limited Data

This study combines custom rules with multiple deep learning predictive models to develop a program that rapidly outputs optimal window configurations meeting lighting requirements based on input building features. The designed program allows for the quick generation of optimal window configurations after reading 2D building layout information. It addresses the current issue in most automatic building layout generation models, where the generated layouts lack window information, making them unsuitable for practical design and diminishing their reference value in terms of building performance. The challenges encountered in this study include how to quickly collect a sufficient number of high-precision building layout datasets that meet the requirements using professional lighting simulation software, and how to train predictive models capable of accurately predicting window configurations under highly complex architectural conditions with a small dataset. To address these issues, this study created a program on the Grasshopper platform (a visual programming plugin for the 3D modeling software Rhino) to perform automated lighting simulations and obtain highly accurate simulation datasets. At the same time, by combining custom rules with multiple deep learning predictive models, the program achieves accurate predictions of window configurations under complex architectural conditions, rather than relying on a single predictive model. This approach significantly improves prediction accuracy and demonstrates the ability to handle complex architectural conditions, even with a small dataset.