Predicting Off‐Fault Deformation From Experimental Strike‐Slip Fault Images Using Convolutional Neural Networks

• Published in: Geophysical research letters 2022-01, Vol.49 (2), p.n/a
• Main Authors: Chaipornkaew, L., Elston, H., Cooke, M., Mukerji, T., Graham, S. A.
• Format: Article
• Language: English
• Subjects: analog faults; Artificial neural networks; Boundary conditions; convolutional neural network (CNN); Crustal deformation; Deformation; Estimates; Experiments; Fault lines; Faults; Geological faults; Learning algorithms; Loading rate; Machine learning; Neural networks; Performance prediction; Plate tectonics; Plates (tectonics); scaled physical experiments; Slip; strike‐slip fault
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2021GL096854

Crustal deformation occurs both as localized slip along faults and distributed deformation off of faults. While there are few robust estimates of off‐fault deformation in nature, scaled physical experiments simulating crustal strike‐slip faulting allow direct measurement of the ratio of fault slip to regional deformation, quantified as kinematic efficiency (KE). We offer an approach to predict KE using a 2D convolutional neural network (CNN) trained directly on fault maps produced by physical experiments. Experiments with different loading rates and basal boundary conditions generate the fault maps throughout the evolution of strike‐slip faults. Strain maps allow us to directly calculate KE and its uncertainty, utilized in the loss function and performance metric. The trained CNN achieves 91% custom accuracy in the KE prediction of an unseen data set. Although the CNN model is trained on scaled experiments, it can predict off‐fault deformation of crustal faults that matches available geologic estimates. Plain Language Summary Where the earth deforms at the boundaries between tectonic plates, some of the deformation is taken up as localized slip along fault surfaces and some of the deformation is distributed around the faults. This distributed deformation is very hard to measure in the crust. To get around this challenge, we create faults in the laboratory and use the direct measurements of the distributed deformation off of faults to train a machine learning model. The trained model performs well at predicting distributed off‐fault deformation from the active fault geometry. Key Points Proposed convolutional neural networks can predict off‐fault deformation directly from binary fault maps Analog models provide abundant and diverse fault map data and calculated deformation labels, an ideal labeled data set for machine learning The trained model achieves 91% accuracy on experimental faults and shows promise in predicting the kinematic efficiency of crustal faults