Reduced Order Models for geotechnical predictions

Finite Element (FE) modelling is routine in geotechnical engineering design and analysis; however, high-fidelity 3D models remain the preserve of highly complex or prestigious geotechnical construction projects, due to their computational cost. A sensitivity study using a complex model can be expensive and time-consuming even using modern hardware. Reduced Order Models (ROMs) offer a solution at potentially vastly lower cost. They are also numerical models of geotechnical problems, where the inputs are a series of defining parameters (geometry, soil properties etc.) and the output is a prediction of interest, e.g. the maximum deflection of a retaining structure. A ROM is developed by analysis of a number of "surrogates" or "snapshots" which are selected instances of results obtained from complex FE modelling. The surrogates are run only once but the ROM can be used to explore predictions in the parameter space in-between the surrogates. In this paper we explain the development of a ROM using a simple geotechnics problem as an example, where advanced neural operator approaches are advantageous. The framework is currently being developed to analyse the behaviour of large 3D braced excavations where the outputs are wall displacements and prop loads, with a particular emphasis on behaviour at the corners.