Oral Presentation

The next decade will likely see an unprecedented increase in indirect, surface observations whose integration with modeling efforts has the potential to transform our understanding of earthquakes. Machine learning/artificial intelligence (ML/AI) approaches for making sense of large data sets are seeing rapid growth in seismology, but an understanding of earthquake processes requires a grounding in physics. In this talk I will provide an overview of a new deep learning technique known as the physics-informed neural network (PINN), which is a data-driven ML method that produces model outcomes constrained by physical laws. I will spotlight recent scientific studies based on PINN models (including new, albeit limited, theoretical advances), and discuss the limitations of PINNs, their possibilities, and many areas ripe for exploration. In particular I will focus on inverse problems, highlighting some of our recent efforts to use PINNs to infer depth-dependency of rate-and-state friction (RSF) parameters using surface measurements and models for elastic/brittle behavior of the solid Earth, which has direct implications for assessing seismic hazard. I will detail how the loss function combines terms from the differential equations governing wave propagation and state evolution, where initial and boundary conditions (including RSF) define the data on which a multi-network PINN is trained. We follow a tiered approach in order to ensure confidence in model outcomes, which follows naturally from decades of earthquake model verification studies in seismology. From here I will outline future research directions, including applications of PINNs to well-instrumented laboratory studies and to surface data from fault zones.