On machine-learning based surrogate modeling of soil-structure interaction problems
Loading...
Authors
Taghavi Ganji, Hamid
Issue Date
2025
Type
Dissertation
Language
en_US
Keywords
Alternative Title
Abstract
This dissertation explores the use of machine learning approaches as surrogate models for problems involving soil-structure interaction. As a first step, we investigate the potential of data-driven models by employing long short-term memory (LSTM) networks to perform uncertainty quantification for a buried structure and predict its seismic response under various earthquake excitations. To improve performance, we enhance conventional LSTM models to accept both time-dependent inputs and scalar random variables. This extension enables the incorporation of a broader range of uncertainties, including mechanical properties such as the soil’s shear wave velocity profile and geometric factors like the structure’s burial depth, in addition to the aleatory uncertainty associated with seismic motions.Next, we employ physics-informed neural networks (PINNs) to model site response analysis in one-dimensional soil columns. For this purpose, we incorporate a viscous damping term in the governing equations using the Kelvin-Voigt damping model. Additionally, we propose a methodology for deriving the dimensionless form of the governing equations under various excitation scenarios. This formulation enables the analysis to be independent of specific mechanical properties and geometric scales. Notably, the dimensionless representation also helps eliminate the need for additional
techniques to address challenges related to weighting loss terms and capturing high-frequency response components. As a final step, we integrate temporal domain decomposition with transfer learning to assess PINN’s performance in modeling site response under long-duration earthquake ground motions.
The acceptable performance of PINNs in predicting the response of one-dimensional systems motivates their extension to two-dimensional problems, where the objective is to predict the seismic response of a near-field domain subjected to remote excitation. To achieve this, we propose a two-step methodology that reformulates the strong-form governing equation in the original problem, with a distant excitation source, into an equivalent problem where the excitation is applied at an auxiliary boundary located near the region of interest. We then evaluate the effectiveness of this approach under two types of truncated boundary conditions: absorbing and transmitting boundary conditions. Consistent with the one-dimensional case, we demonstrate that the use of dimensionless governing equations yields acceptable results, regardless of the specific problem configuration.