Transfer Learning From Numerical Experiments Predict Fault Friction in the Laboratory

Data-driven machine-learning for predicting instantaneous and future fault-slip characteristics in laboratory experiments has recently progressed markedly, with a large number of papers published on the topic as well as a Kaggle earthquake prediction competition sponsored by the US Department of Energy Office of Science and Kaggle. The progress is due primarily to the availability of large training data sets available from some laboratory experiments. In Earth however, earthquake interevent times range from 10s-100s of years, whereas geophysical data typically exist for only a portion of an earthquake cycle. Sparse data present a serious challenge to training machine learning models if the goal is predicting slip on seismogenic faults in Earth. Here, we describe a prototype transfer learning approach developed using fault-slip numerical simulations and laboratory data to address the problem of sparse data. Numerical simulation data are used to train a convolutional encoder-decoder that predicts fault-slip behavior in the laboratory experiments. The model learns a mapping between acoustic emission histories and fault friction from numerical simulations, and generalizes to produce accurate predictions of laboratory fault friction. The encoder-decoder model latent space is then trained further using only a portion of data from a single laboratory earthquake-cycle, as an analog to a fault in Earth where data is sparse. Notably, predictions markedly improve by further training the model latent space. The transfer learning results elucidate the potential of using machine learning models trained on numerical simulations and fine-tuned with small geophysical data sets for potential applications to faults in Earth.