Generative model to improve the physical simulation of normalized seismic acceleration

Abstract

Earthquake physics-based simulations have revealed their accuracy limit in high-fidelity broadband strong ground motion scenario prediction. Earlier attempts leveraging AI tried to fill the gap between the low-frequency high-fidelity numerical simulation and the high-frequency accuracy and realism demanded in earthquake engineering. Our research is built on top of the SeismoALICE solution proposed by Gatti and Clouteau (2020), an AI generative approach that renders broadband (0-20 Hz) single-station accelerograms conditioned by low-frequency physics-based simulation outcomes. First, we developed a novel neural architecture based on Encoder, Decoder, and Discriminators that integrates Conformer's advanced attention techniques (Gulati et al., 2020), stabilising the training and producing realistic output for the generation. This novel architecture is trained according to Adversarial Learning Inference with Conditional Entropy (ALICE, Li et al., 2017). Second, our approach employs a similarity evaluation technique called Hyper-Spherical Loss (HSL) in the time domain and an adaptation of the Focal Frequency Loss (FFL) for time series. Our investigation demonstrates that Conformer architecture outperforms previous approaches in super-resolution of earthquake data, such as the STanford EArthquake Dataset (STEAD) Dataset (Mousavi et al., 2019). We finally showcase the enhanced strong motion synthesizer to predict the seismic response at the Cruas Nuclear Power Plant during the 2019 MW 4.9 Le Teil earthquake, providing insightful perspective for future large-scale applications as a downstream generative pipeline.