Physics-guided machine learning for laboratory earthquake prediction

Recent laboratory studies of fault friction have shown that deep learning can accurately predict the magnitude and timing of stick-slip sliding events, the laboratory equivalent of earthquakes, from the preceding acoustic emissions (AE) events or time-lapse active-source ultrasonic signals. While there are observations that provide insight into the physics of these predictions, the underlying precursory mechanisms are not fully understood. Furthermore, these purely data-driven models require a large amount of training data and may not generalize well outside their training domain. Here, we present a physics-guided machine learning approach - by incorporating the relevant physics directly in the prediction model architecture - with the objectives of enhancing model predictions and generalizability as well as reducing the amount of required training data. We use data from well-controlled double-direct shear laboratory friction experiments on Westerly granite blocks exhibiting numerous regular and irregular stick-slip cycles. Simultaneously, AEs are recorded while the faults are also regularly probed by ultrasonic waves transmitted through the fault zone to monitor the evolution of the contact stiffness during shearing. Our physics-guided ML models take features extracted from AE time series or time-lapse active source ultrasonic signals and predict the shear stress history, which gives both the timing and size of the laboratory earthquakes. The models are constrained by friction laws as well as simplified physical laws governing ultrasonic transmission and AE generation. Our findings indicate that physics-guided ML models outperform purely data-driven models in important ways; they provide accurate predictions even with little training data and transfer learning is greatly enhanced when physics constraints are incorporated. These findings have important implications for earthquake predictions in the field, where training data are scarce.