SCEC Award Number 22138 View PDF
Proposal Category Collaborative Proposal (Integration and Theory)
Proposal Title Earthquake Gates: Using machine learning to characterize the branch-fault rupture propagation along the Cajon Pass in Southern California from multi-cycle simulations and realistic fault geometry
Investigator(s)
Name Organization
Abhijit Ghosh University of California, Riverside David Oglesby University of California, Riverside Roby Douilly University of California, Riverside Eric Daub Alan Turing Institute
Other Participants
SCEC Priorities 1d, 4a, 1e SCEC Groups FARM, SAFS
Report Due Date 03/15/2023 Date Report Submitted 12/05/2023
Project Abstract
 Anticipating potential earthquake rupture paths in the San Andreas fault (SAF) system poses considerable challenges due to its complex geometry, stress conditions, and frictional properties. A commonly used approach relies on iterative simulations to understand the parameters determining rupture propagation. Such methods are inefficient because they do not cover every scenario for the range of parameters involved and are computationally expensive. In this study, we use machine learning (ML) algorithms trained from simulation outputs to better understand the physical parameters that are most important in controlling rupture propagation at geometrical complexities.
We show how machine learning models employing gradient boosting (xgboost) and random forest (scikit-learn) can be used for classification of rupture at the junction of the Garlock and SAF system. The training and testing data for these models have been processed from a database of ~9800 rupture simulations generated from Rate-State earthquake simulator (RSQSim), reflecting variations in stress conditions and frictional parameters.
We show that the models quantitatively identify different combinations of critical parameters which affect the rupture propagation at the branch, and are able to identify important features which are relatively important using the feature importance tools of xgboost. We demonstrate that for certain scenarios the ML models demonstrate considerable testing accuracy (>65%) in classifying the propagation direction and augmentation of features in the input training data leads to improvements in the accuracy both in terms of precision and recall.
Intellectual Merit An important factor in assessing seismic hazard is trying to understand how rupture will behave at geometric complexities along the fault system. Many large earthquakes have occurred on complex branch fault systems and exhibited a variety of behaviors. Sometimes the rupture is constrained to the nucleated fault and does not rupture connecting faults; such was the case for the 2010 Haiti Earthquake (Calais et al., 2010; Douilly et al., 2013). Other times the rupture completely transfers from one fault onto another. During the 2002 Denali Earthquake, upon reaching the branch intersection the rupture fully transferred from the Denali fault to the Totschunda fault, with no measurable slip on the Denali fault past the intersection (Eberhart-Phillips et al., 2003; Haeussler et al., 2004; Ratchkovski et al., 2004). Occasionally the rupture propagates through the intersection rupturing multiple fault segments, examples include the 1979 Imperial Valley Earthquake (Archuleta, 1982, 1984) and notably the recent 2023 Türkiye earthquake sequence (Barbot et al., 2023; Melgar et al., 2023). In southern California there are numerous branch fault systems, and it is important to understand how earthquake rupture behaves near such brunches, and what controls its propagation. In this study we focus on the intersection of the San Andreas and Garlock faults, two of the longest faults in California. This study is aligned with SCEC’s interest to study San Andreas Fault system, fault rupture and understand stress dynamics in complex realistic fault geometry.
Broader Impacts This project helps in training two graduate students who worked on this project. It fosters collaboration between three earthquake scientists focusing on different geophysical tools. This project is closely related to seismic hazards and contribute to our understanding of earthquake risks in California. In addition, results of this projects can be used by researchers in other fields to inform their studies and interpretations.
Exemplary Figure Fig 2: Flowchart of steps involved for the generation of machine learning classification data from the simulation outputs.