Shear Failure Prediction Using a Physics-guided Deep Learning Framework

Prediction of micro-earthquakes when monitoring geothermal reservoirs, CO2 storage sites, and unconventional reservoirs using active source seismic data although challenging, is essential to ensuring the safety of the operations. Laboratory-scale friction experiments have shown that changes in elastic wave amplitude and speed during active source monitoring carry precursory information about the upcoming failure. The data-driven deep learning models predict the timing and size (shear stress drop) of the laboratory earthquakes with great accuracy using the elastic wave attributes. However, data-driven models require training on large datasets and overlook the physical laws controlling the earthquake rupture. As such, the model may perform well for a particular dataset but fail to provide satisfactory predictions for a different albeit closely related dataset. To incorporate the domain knowledge and address the model transferability challenge, in this study, a physics-guided deep learning approach is implemented to forecast the earthquake timing and size. We demonstrate the efficacy of this approach using data from a laboratory-scale friction experiment conducted close to the stability boundary producing numerous regular and irregular seismic cycles. A pair of P-wave ultrasonic transducers is used to probe the laboratory fault throughout the seismic cycles. Two sets of features are extracted from the ultrasonic data: (1) physics-based wave speed and amplitude, and (2) automatically extracted features from recorded waveforms using Convolutional Neural Network (CNN) and Scattering Network algorithm. The data-driven predictions obtained using Long Short-term Memory (LSTM) and Multilayer Perceptron (MLP) models are used as a baseline. Besides time to failure and shear stress, fault slip rate is also predicted and used in the physical constraint formulations. The rate-and-state friction law and elastic coupling relation are integrated into the LSTM and MLP model architectures to modify the loss function. We compare the data-driven and physics-guided deep learning model performances for both sets of features. A physics-guided framework can improve model generalizability expected to improved earthquake and micro-seismicity predictions informed by the physics laws dictating the fault state