← Back to Sessions

Predicting Off-Fault Deformation Using Convolutional Neural Networks Trained on Experimental Strike-Slip Faults

Description: 

Surface offsets might not represent slip at seismogenic depths because ruptures produce shallow distributed off-fault deformation. Scaled physical experiments simulate off-fault deformation processes and provide direct observations of deformation partitioning during fault evolution. Using experimental fault maps, we train and test a 2D Convolutional Neural Network (CNN) that can predict off-fault deformation of strike-slip fault trace maps. The CNN predicts kinematic efficiency (KE), the ratio of strike-slip rate along the faults to the total velocity. High kinematic efficiency indicates mature and through-going fault surfaces that require less work to maintain deformation resulting in small off-fault deformation and limited shallow slip deficit. On the other hand, immature strike-slip faults with segmented and complex trace geometry produce greater off-fault deformation and larger shallow slip deficit.

We train the CNN on experimental strike-slip faults in both kaolin and sand with different loading rates and basal conditions to simulate a wide range of conditions that control evolution of fault geometry and off-fault deformation. The CNN hyperparameters minimize our custom loss function. We compare how the CNN trained on faults that grow in one rheology and boundary condition perform in predicting KE of unseen fault maps with different conditions to assess the predictive power of the CNN. All experimental fault maps are combined for training the CNN that incorporates expected variations in the crust and tested on crustal strike-slip fault maps of different maturity in southern California. This study shows the potential that deep learning trained on experimentally produced faults has in shedding insight into the strain partitioning of crustal faults.

Session: From Earthquakes to Plate Boundaries: Insights Into Fault Behavior Spanning Seconds to Millennia

Type: Oral

Date: 4/20/2023

Presentation Time: 10:30 AM (local time)

Presenting Author: Christ F. Ramos Sanchez

Student Presenter: Yes

Invited Presentation: 

Authors