High-Resolution Inversion of Seismic Stress Fields and Dynamic Hazard Analysis via Adaptive Physics-Informed Neural Networks and NGL GNSS Data Fusion

Abstract

Proper inversion of the seismic stress field is essential in the studies of fault activity as well as earthquake hazards. The simplified physical models and paucity of observations constrain the traditional methods. The proposed paper offers a new framework, where ANN is Adaptive Physics-Informed Neural Network (Adaptive-PINN), which combines high-precision GNSS surface velocity field values in the Nevada Geodetic Laboratory (NGL) and in its result, achieves high-resolution inversion of the crustal stress field and dynamic hazard analysis. It uses deep neural networks to build a nonlinear model of geographic coordinates to stress-field parameters and applies physical plausibility through embedded constraints (e.g., stress equilibrium equations and constitutive relations). The adaptive mechanism that is proposed uses the weights of the losses dynamically in such a way that they place more emphasis on regions with large physical residuals, which has a substantial effect on the interpretability and convergence ability. An example of a case study on how global GNSS velocity fields can be used to create spatially continuous, high-resolution stress maps and an earthquake hazard index. Conclusions indicate that the technique is effective to not only recreate stress concentrations along established active tectonic belts but also possible high risk zones which has created a potent new resource in the process of earthquake dynamics, as well as disaster risk assessment.