CAREER: Physics-Informed Deep Learning for Understanding Earthquake Slip Complexity

What it is about one fault that causes it to slip suddenly, unleashing catastrophic earthquakes, while another just creeps along steadily or produces smaller, more frequent earthquakes? This is difficult to assess because faults cannot be directly observed at depths where earthquakes start, typically 5 to 15 miles below ground. We must rely instead on indirect measurements made by instruments at the Earth's surface, and computer models representing the fault and how it slips in response to pressures deep in the Earth. Properties of the virtual fault and surrounding rock can be repeatedly adjusted until the model outputs data that closely match real-world observations from seismometers and other instruments. This process is slow and expensive, even when scientists use clever strategies. Dr. Erickson and her group will see whether a new artificial intelligence scheme called a "Physics-Informed Neural Network" (PINN) can learn how to efficiently adjust fault model properties to rapidly fit observational data. They will test their PINN first on data from laboratory fault experiments to see how it performs at estimating the already-known fault properties, and then train it until it learns to do this well. Then they will apply the PINN to data from the Pacific Northwest and Costa Rica, where properties and physics of dangerous offshore faults need to be better understood. In addition to their main project, Erickson's team will lead short courses on modern computer programming, data analysis, and AI methods for community college students, using datasets and techniques from this project. Dr. Erickson and her group will apply a deep learning learning technique called the Physics-Informed Neural Network (PINN) to study fault slip, using synthetic and laboratory data, as well as geodetic and seismic data from the Cascadia and Costa Rica subduction zones. Scientific questions concern how heterogeneous fault friction and material properties in subduction zone settings affect fault zone slip, stress, and pore pressure; and how/whether PINNs can be applied to studies of this kind. PINN-based solutions for slip, stress, and pore pressure will be compared with those from traditional computational methods to verify the PINN-based solutions and assess their computational advantages and limitations. The three thrusts of the project are (1) developing the theoretical and computational framework; (2) verifying, validating, and applying methods to (i) analytical solutions and community code verification exercises, (ii) controlled laboratory fault slip experiments, and (iii) natural faults; and (3) training and mentoring students. This project will support two-week mini research experiences for ten community college students, multidisciplinary training at UO and two other universities for several graduate students, and an international collaboration with scientists from Costa Rica. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.