Prediction of Off-Fault Deformation from Strike-slip Fault Structures in clay and sand experiments using Convolutional Neural Networks

Crustal deformation occurs both as localized slip along faults and distributed deformation off faults; however, we have few robust geologic estimates of off-fault deformation over multiple earthquake cycles. Scaled physical experiments simulate crustal strike-slip faulting and allow direct measurement of the ratio of fault slip to regional deformation, quantified as Kinematic Efficiency (KE). We offer an approach for KE prediction using a 2D Convolutional Neural Network (CNN) trained directly on images of fault maps produced by physical experiments of strike-slip loading of wet kaolin. A suite of experiments with different loading rate and basal boundary conditions, contribute over 13,000 fault maps throughout strike-slip fault evolution. Strain maps allow us to directly calculate KE and its uncertainty, utilized in the loss function and performance metric. The trained CNN achieves 91% accuracy in KE prediction of an unseen dataset. We then apply this CNN trained on wet kaolin experiments to strike-slip experiments in dry sand. The different rheology of sand and kaolin may lead to different relationships between fault geometry and off-fault deformation, which can be detected by differences in the predictive power of the CNN trained only on kaolin.  We also apply the trained CNN to crustal maps of off-fault deformation over coseismic, 10ka and 1 Ma time scales. The CNN predicted off-fault deformation overlap available geologic estimates.