Fast simulation of the deformed foundation pit using graph neural networks

Abstract Foundation pits are widely distributed in modern cities, the construction of which are constrained by complex geological conditions and surrounding infrastructures. Ability to reliably control deformations and surface settlements is essential to the safety and durability of geotechnical structures, particularly in densely populated regions. There is a strong need to connect the monitoring data with fast interpretations from numerical modeling. This work aims at developing an approach to predict the deformation of foundation pits by integrating rapid simulation from the Graph Neural Network (GNN) and measurements from real engineering projects. First, predictions from 2D GNN models of deformed foundations were compared with results from Pix2Pix. A good agreement was obtained for deformation predictions between graph-based (GNN) and image-based (Pix2Pix) machine learning models. Subsequently, 3D GNN models were developed to accommodate 3D geometries, multiple materials, and complex interfaces. Good agreement was obtained for predicted deformation distributions between GNN and finite element simulations for a variety of materials and structures. For the same level of accuracy, the computational cost was reduced from 40 s in finite element method to less than 0.05 s in GNN model. Finally, the GNN model was used to predict deformation of an actual foundation pit at different time stages. Results showed that predictions from the GNN model agreed well with the actual monitoring data, from which the essential materials parameters were extracted along the transient deformation process.