Automated simulation modeling and sustainable performance evaluation for flat layout design

Generative design is increasingly applied to flat layouts. However, generative models lack automated and efficient methods for performance evaluation, including (1) automatic conversion of flat layout images into simulation models and (2) efficient performance evaluation at the flat level. This paper addresses these issues by proposing an automated image-based performance evaluation surrogate model. Firstly, a new geometric feature set is proposed. Secondly, the Image2Sim algorithm is developed for automated flat layout modeling from images in the RPLAN dataset. Finally, a graph-aware extreme gradient boosting (GAXGBoost) surrogate model is developed for flat-level performance evaluation. Results demonstrate that (1) the Image2Sim algorithm reduces the failure rate by 8 % and simulation modeling time from 333 days to 2 days; (2) the GAXGBoost outperforms XGBoost, MARS, GNN, and ANN across all metrics. The GAXGBoost provides accurate and timely feedback on flat layout performance, thus facilitating performance-driven generative design in the early design stage.