[S-CG50] Driving Solid Earth Science through Machine Learning

[SCG50-P04] Empirical knowledge-informed deep learning approach for ground motion prediction equations

*Tomohisa Okazaki¹, Asako Iwaki², Nobuyuki Morikawa², Hiroyuki Fujiwara², Naonori Ueda¹ (1.RIKEN Center for Advanced Intelligence Project, 2.National Research Institute for Earth Science and Disaster Resilience)

Physics-informed neural networks (PINNs) can solve differential equations by incorporating their residuals into the loss functions of neural networks (NNs) as a soft constraint. The knowledge on natural phenomena ranges from differential equations, symmetries, scaling laws, to empirical relations. Such domain knowledge can introduce a physical regularization to NNs through the loss functions. In particular, the soft nature of constraints would be suitable for ambiguous empirical relations.

Here, we investigate this approach for ground motion prediction equations. Data-driven NN models show high accuracy in data-rich parameter domains, whereas can predict unrealistic values in extrapolation domains. For example, we observed overestimation of near-field acceleration response spectra in long periods predicted by a monotonic NN model. By incorporating empirical knowledge on ground motions, we confirmed that such overestimation was suppressed, whereas estimations in data-rich domain were not significantly biased. We also found that predictions in extrapolation domains are sensitive to the form of knowledge and regularization parameter values.