A mechanism-informed neural network with a physical intermediate layer for predicting wall deflection induced by braced excavations in soft soil

Abstract

Accurate advance prediction of wall deflection induced by braced excavations is crucial for preventing potential damage to retaining structures and the surrounding environment. Although previous studies have utilized spatiotemporal deep learning models for this purpose, they have overlooked the underlying physical mechanisms of wall deflection. To address this limitation and enhance the interpretability and transferability of deep learning models, this paper proposes a mechanism-informed neural network for predicting wall deflection, where the physical mechanism based on the beam on elastic foundation method is hardcoded in the neural network architecture. In the proposed model, a physical intermediate layer is designed to mimic the effects of the horizontal load behind the wall, and a monotonicity-preserving long short-term memory network is devised to capture the inherent monotonic characteristics of the horizontal load. Additionally, a hybrid loss function is designed to simultaneously constrain the outputs of both the final layer and the physical intermediate layer. The performance of the proposed model was validated by different excavation projects, with its significant superiority over baseline models demonstrated. The proposed model demonstrates a strong capability and generalizability for accurately forecasting wall deflections in advance by incorporating the physical mechanism.