Predicting fault slip via transfer learning

Area of Science:

• Geophysics
• Machine Learning
• Computational Seismology

Background:

• Machine learning models for predicting fault slip show promise but require extensive training data.
• Earthquake data is sparse, posing a significant challenge for training predictive models for real-world fault behavior.

Purpose of the Study:

• To develop a transfer learning approach for predicting fault slip using numerical simulations.
• To investigate the efficacy of training a convolutional encoder-decoder model with simulated data and fine-tuning it with laboratory experimental data.

Main Methods:

• A convolutional encoder-decoder model was trained using acoustic emission and fault friction data from numerical simulations.
• The trained model was then applied to predict fault-slip behavior in laboratory experiments.
• Further fine-tuning of the model's latent space was performed using limited data from a single laboratory earthquake cycle.

Main Results:

• The transfer learning model accurately predicted fault friction in laboratory experiments.
• Fine-tuning the model with a small subset of laboratory data significantly improved prediction accuracy.
• The study demonstrates successful generalization from simulated to experimental data.

Conclusions:

• Transfer learning offers a viable solution for training machine learning models on sparse geophysical data.
• Models trained on numerical simulations can be effectively adapted for predicting fault slip in laboratory settings.
• This approach holds potential for improving earthquake forecasting and understanding fault behavior in Earth.