Machine learning can predict if an earthquake would break through a fault

Sabber Ahamed, Eric Daub

Dynamic rupture propagation is a challenging problem due to uncertainty regarding the underlying physics of earthquake slip, and the stress conditions and frictional properties of fault are not well constrained. These unknown initial stresses and friction combine with fault geometry to control the rupture process and determine the dynamics of slip and the resulting ground motions.

The San Andreas Fault in the Carrizo Plain, about 100 miles northwest of Los Angeles, is seen from the sky in this 2007 photo. (Photo/Ian Kluft). The photo is taken from: https://news.usc.edu/82215/usc-geologist-looks-below-solid-ground-to-study-little-known-faults/

However, because the earthquake rupture problem is highly nonlinear, determining parameter values is often done by making simplifying assumptions combined with trial and error, which computationally and numerically expensive. To improve the ability to determine reasonable friction and stress parameters, we use machine learning methods to develop models to predict if rupture can break through a fault with geometric heterogeneities.

The illustration shows the parameters learned by the ANN for the rupture model. The network has one hidden layer with twelve nodes. The left panel shows the weights that combine hidden and input units. Eight input parameters are on the horizontal scale, and the twelve hidden units are on the vertical scale. All input parameters are normalized to have zero mean and a standard deviation of unity. The colors in each row indicate how the parameters are combined to form each hidden unit. The right panel shows that the weights combine in the hidden units into the single output unit. A substantial positive value of the output unit indicates that the rupture is predicted to break the barrier, while a large negative amount of the output unit suggests that the rupture is not anticipated to break the barrier. That allows a physical understanding of the parameters selected by the neural network.

We create two models using the artificial neural network (ANN), and the random forest decision tree (RFD) algorithms. We train the models using a database of 1000 dynamic rupture simulations with varying fault geometry, stress conditions, and friction parameters. We also rigorously validate and test the predictive power of the models using additional simulations. Both RFD and ANN models can predict if a rupture can break through the geometric complexity with 82% accuracy on our test simulations, and require significantly fewer computational resources to predict rupture characteristics once the models have been trained.

The model parameters that are determined through machine learning are also useful in determining what the most important physical parameters that control rupture propagation, providing new insights into the highly nonlinear process of fault rupture.

Both of the models are computationally efficient such that the 400 testings took a fraction of a second, leading to potential applications of dynamic rupture that have previously not been possible due to the computational demands of simulations.

The details of the code can be found in Github. A poster of this project can be downloaded from this link. The details research findings will be published in Journal of geophysical research which is under review right now. As soon as the paper get published i will add a link of the paper.