Learning a potential formulation for rate-and-state friction

Empirical rate-and-state friction laws are widely used in geophysics and engineering to simulate interface slip. They postulate that the friction coefficient depends on the local slip rate and a state variable that reflects the history of slip. Depending on the parameters, rate-and-state friction can be either rate-strengthening, leading to steady slip, or rate-weakening, leading to unsteady stick-slip behavior modeling earthquakes. Rate-and-state friction does not have a potential or variational formulation, making implicit solution approaches difficult and implementation numerically expensive. In this work, we propose a potential formulation for the rate-and-state friction. We formulate the potentials as neural networks and train them so that the resulting behavior emulates the empirical rate-and-state friction. We show that this potential formulation enables implicit time discretization leading to efficient numerical implementation.