Seismic or aseismic: a machine learning approach to classification of frictional fault systems

Abstract:

Spring-slider models of frictional slip, based on rate-and-state laws, provide insights into the dynamics of earthquake rupture. In particular, they determine whether a fault slips seismically or aseismically. The stability of frictional slip in these models depends on the evolution of the frictional resistance, which in turn depends on several fault parameters—the most common and widely accepted model being the rate-and-state model of friction. In the case of injection-induced seismicity, a recent extension of these models to account for poroelastic effects showed that the seismic response also depends on the injection rate and hydraulic diffusivity of the reservoir rock surrounding the fault [1]. The motion of the poroelastic spring-slider model with an evolving pore pressure depends on the solution of 4 coupled nonlinear Ordinary Differential Equations (ODEs) and 8 dimensionless parameters describing fault and reservoir properties. While it is conceivable to fully characterize the parameter space for this single degree-of-freedom model, this is certainly intractable for complex models.

In this study, we explore the use of machine learning to construct a mapping from the parameter space to event classification for the rate-and-state friction model considering the poroelastic coupling. Taking advantage of the relatively fast solution of the system of ODEs for the poroelastic spring-slider model, we first explore the construction of such a mapping using a purely data-driven approach by drawing many realizations. Next, we explore the use of a recent paradigm of machine learning, Physics-Informed Neural Networks (PINNs), to construct this mapping. In this approach, we incorporate the governing equations as constraints in the loss function of the deep neural network [2]. We assess the increased robustness and computational efficiency of the PINN classifier versus the data-driven classifier, but also point to the challenges posed by the extreme disparity in time scales between the aseismic and seismic stages of the slip dynamics. We believe that this proof-of-concept study provides a new paradigm for the classification of induced-seismicity hazard in three-dimensional fault systems.

[1] M. Alghannam, and R. Juanes. Understanding rate effects in injection-induced earthquakes.Nature Comm., 11, 3053 (2020).

[2] E. Haghighat, and R. Juanes. SciANN: A Keras/TensorFlow wrapper for scientific computations and physics-informed deep learning using artificial neural networks. arXiv:2005.08803 (2020).