Application of Neural Networks for Estimating Coseismic Slip Distribution Using Synthetic GNSS Data

Accurately and rapidly estimating coseismic slip is crucial for characterizing the rupture and magnitude of earthquakes and improving early warning systems. However, slip inversions often involve hyperparameters related to prior information, whose selection significantly affects solution efficiency and quality. In this study, we present a novel method for estimating fault slip distributions from GNSS displacements using neural networks. The method was initially developed with synthetic models to define an optimal architecture for accurately recovering target slip. This approach demonstrated exceptional computational efficiency, delivering accurate slip distribution predictions in just 0.07 seconds without specialized hardware. The method was then validated with real-world data from the 2015 Mw 8.3 Illapel earthquake, achieving a GNSS displacement RMSE of 0.07 m and yielding a slip distribution consistent with published solutions. Compared to Regularized Least Squares (RLS) inversion, the neural network estimated slip closer to the trench, aligning with tsunami observations despite slightly higher residuals. Additionally, hyperparameter exploration revealed that using the GELU activation function and a 35% dropout rate provided the best balance. Model performance improved with larger datasets, and while reducing GNSS stations increased uncertainty, more data enhanced accuracy. These findings highlight the importance of hyperparameter tuning and data selection in improving slip estimations, offering insights for future improvements.