Reining‐In the Spring‐Slider With Reinforcement Learning

Abstract Subsurface fluids are important to earthquake physics since they influence every phase of the earthquake cycle: from inducing earthquakes, generating slow slip, dynamically weakening a fault, to producing afterslip. Despite this prominent role, comparatively little thought has been directed toward intentionally controlling fault slip. I take the spring‐slider as the simplest analogue for earthquake‐like motion and train a deep reinforcement learning agent to design fluid injection that reins‐in slip motion (i.e., controls slip velocity). These reining algorithms can mitigate stick‐slip instability via a three‐step injection policy. First, by injecting to induce slip nucleation; second, by harnessed withdrawal that governs slip speed; third, by injection‐driven steady‐state sliding. These numerical simulations are supported by theoretical derivations that show fault slip acceleration can be reined‐in by balancing pressurization rate with state evolution changes. I discuss the relevance to prior studies, robustness of the algorithms, and discuss potential limitations/solutions to scaled‐up problems. Together, these results suggest that spring‐sliders could be tamed with a carefully designed injection policy.