Development and Verification of a Neural Operator Based Nonlinear Site Response Module for the San Francisco Bay Area

Domniki Asimaki, Flora Xia, & Grigorios Lavrentiadis

Published September 8, 2024, SCEC Contribution #14059, 2024 SCEC Annual Meeting Poster #182

The proliferation of seismic record repositories through the expansion of seismic networks and high-performance computer simulations of earthquake scenarios, have produced ground-motion databases rich enough to enable data-driven models of ground-motion synthesis that account for complex site effects. Data-driven methods offer a novel approach to describing these processes by directly learning the governing laws from sufficiently rich training data, while avoiding the use of simplified assumptions that limit the realism of models developed with traditional statistical tools. In this study, we demonstrate this new paradigm of learning the underlying physics in a data-driven framework consisting of a Fourier Neural Operator (FNO), which modifies outcrop ground motions to account for the full nonlinear response of the near-surface soil layers. The FNO model was trained on non-linear one-dimensional wave propagation through smooth Bay Area velocity profiles using the site-response software, PySeismoSoil, and ground motions from the SCEC Broadband Platform (BBP). A key advantage of the neural operator architecture of FNO compared to traditional neural networks, is its ability to learn the mapping between continuous function spaces as opposed to finite-dimensional sets, rendering the training and application of the model resolutions invariant. In other words, training the model can include input signals of different sampling frequencies without loss of information or generation of artifacts, while prediction can be performed on sampling frequencies independent of training. Verification analyses through residual and goodness of fit evaluations demonstrate that FNO can correctly estimate the non-linear amplification for ground motions and profiles not included in the training dataset for the 0.1 to 30hz frequency range. By appropriately conditioning data-driven algorithms, our work demonstrates the potential of using these methods to learn increasingly complex physics and their uncertainty over the entire frequency range of engineering interest, and to generate on demand time-histories appropriate for engineering design with a high degree of realism.

Citation
Asimaki, D., Xia, F., & Lavrentiadis, G. (2024, 09). Development and Verification of a Neural Operator Based Nonlinear Site Response Module for the San Francisco Bay Area. Poster Presentation at 2024 SCEC Annual Meeting.

Related Projects & Working Groups
Ground Motions (GM)