Prediction of off-fault deformation from experimental fault structure using convolutional neural networks

Geodetic measurements record contributions of both deep loading along active faults and distributed deformation of the rock between faults. However, the latter has few robust estimates. My research offers an alternative off-fault estimation using 2D Convolutional Neural Network that is trained directly using physical experiments scaled to crustal strike-slip fault development. Direct and detailed observation of fault traces suggest smoother traces can more efficiently accommodate strike-slip faults (measured as Kinematic Efficiency, KE) compared to complex traces. A variety of experiments, representing ranges of ductile and brittle deformation, contribute to over 10,000 input images prior to further augmentation.  After iterating over hyperparameter selections and model architectures, we are able to achieve >90% accuracy in KE prediction. By comparing the performance of different subsets of data we can learn what aspects of fault development have greater impact on the relationship between fault pattern and off-fault deformation. We are expanding the datasets to include a wider range of analog material rheology (linear elastic, elastic perfectly plastic and viscoelastic) and material thickness. We aim to produce a regional off-fault deformation heat map from strike-slip fault traces and highlight the regional deformation due to significant off-fault deformation.