Multivariate fusion-based surrogate modeling for predicting excavation-induced full-field vertical soil displacement

Abstract

In geotechnical engineering, foundation pit projects face growing challenges in predicting excavation-induced vertical soil displacement under complex and variable conditions. This paper presents PitGAN, a generalizable surrogate modeling approach designed to deliver accurate, full-field predictions of vertical soil displacement for field-scale risk assessment. PitGAN integrates (1) region-adaptive data generation using hydro-mechanical coupled simulations, (2) multivariate fusion preprocessing guided by 2D spatial knowledge, and (3) generative adversarial network-based learning with attention mechanisms. A region-specific PitGAN model developed for Shanghai, China achieved accuracy comparable to numerical simulations while identifying interpretable geotechnical patterns and providing precise displacement predictions in a real-world case with complex excavation conditions. Compared with existing approaches, PitGAN delivers full-field predictions with higher accuracy and improved spatial fidelity in capturing localized peak deformations, while achieving inference speeds over three orders of magnitude faster than numerical methods, thereby making it well suited for real-time, field-scale risk assessment and safety management.