[Paper Review] Deep learning for laboratory earthquake prediction and autoregressive forecasting of fault zone stress

Laura Laurenti, Elisa Tinti+3 more|arXiv (Cornell University)|Mar 24, 2022

Seismology and Earthquake StudiesComputer Science52 references87 citations

TL;DR

This study introduces a novel deep learning (DL) framework for predicting and autoregressively forecasting laboratory earthquake timing and fault zone shear stress using acoustic emission (AE) signals. It demonstrates that LSTM, Temporal Convolutional Network (TCN), and Transformer architectures outperform prior models in predicting time-to-failure and forecasting future stress evolution, with TCN and Transformer showing strong performance on aperiodic and noisy sequences, confirming AE as a reliable stress fingerprint.

ABSTRACT

Earthquake forecasting and prediction have long and in some cases sordid histories but recent work has rekindled interest based on advances in early warning, hazard assessment for induced seismicity and successful prediction of laboratory earthquakes. In the lab, frictional stick-slip events provide an analog for earthquakes and the seismic cycle. Labquakes are ideal targets for machine learning (ML) because they can be produced in long sequences under controlled conditions. Recent works show that ML can predict several aspects of labquakes using fault zone acoustic emissions. Here, we generalize these results and explore deep learning (DL) methods for labquake prediction and autoregressive (AR) forecasting. DL improves existing ML methods of labquake prediction. AR methods allow forecasting at future horizons via iterative predictions. We demonstrate that DL models based on Long-Short Term Memory (LSTM) and Convolution Neural Networks predict labquakes under several conditions, and that fault zone stress can be predicted with fidelity, confirming that acoustic energy is a fingerprint of fault zone stress. We predict also time to start of failure (TTsF) and time to the end of Failure (TTeF) for labquakes. Interestingly, TTeF is successfully predicted in all seismic cycles, while the TTsF prediction varies with the amount of preseismic fault creep. We report AR methods to forecast the evolution of fault stress using three sequence modeling frameworks: LSTM, Temporal Convolution Network and Transformer Network. AR forecasting is distinct from existing predictive models, which predict only a target variable at a specific time. The results for forecasting beyond a single seismic cycle are limited but encouraging. Our ML/DL models outperform the state-of-the-art and our autoregressive model represents a novel framework that could enhance current methods of earthquake forecasting.

Motivation & Objective

• To improve laboratory earthquake prediction accuracy using deep learning on acoustic emission (AE) signals.
• To develop and evaluate autoregressive (AR) forecasting models for predicting future fault zone shear stress from past stress values.
• To test the robustness of DL models across diverse fault conditions, including pre-seismic creep, aperiodic events, and alternating slow/fast slip events.
• To investigate whether acoustic energy is a reliable proxy for fault zone stress, enabling physical insight into seismic cycle dynamics.
• To establish a new framework for time-series forecasting in seismology that extends beyond single-event prediction to multi-step future evolution.

Proposed method

• Employed three deep learning architectures: Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), and Transformer (TF) for sequence modeling of AE and stress data.
• Used autoregressive forecasting by training models to predict future stress values using only past stress values as input, enabling multi-step predictions without external labels.
• Trained models on long sequences of AE and shear stress data from laboratory fault experiments under controlled conditions.
• Preprocessed AE signals by computing variance over sliding windows to extract features related to stress accumulation and failure precursors.
• Applied transfer learning principles by pretraining Transformer models on synthetic sine waves to improve convergence on limited real lab data.
• Evaluated model performance using metrics like mean absolute error (MAE) and correlation coefficient between predicted and actual time-to-failure (TTsF, TTeF) and stress values.

Experimental results

Research questions

• RQ1 Can deep learning models predict the time to start of failure (TTsF) and time to end of failure (TTeF) in laboratory earthquakes using only acoustic emission signals?
• RQ2 Can autoregressive deep learning models forecast the future evolution of fault zone shear stress based on its historical time series?
• RQ3 How do different deep learning architectures (LSTM, TCN, Transformer) compare in predicting and forecasting labquake dynamics under varying fault conditions?
• RQ4 Is acoustic emission variance a reliable proxy for fault zone shear stress, especially during aperiodic or creep-influenced failure cycles?
• RQ5 To what extent can models generalize across different fault zones and failure patterns, including alternating slow and fast events?

Key findings

• The proposed deep learning models significantly outperform state-of-the-art machine learning methods in predicting time-to-failure for laboratory earthquakes.
• Time to End of Failure (TTeF) was successfully predicted in all seismic cycles, while Time to Start of Failure (TTsF) prediction accuracy varied depending on the presence of preseismic creep.
• Autoregressive forecasting using TCN and Transformer models successfully predicted future fault zone shear stress over multiple time steps, representing a novel approach in seismology.
• The TCN model outperformed LSTM in stress forecasting, likely due to better handling of long sequences and periodic stress patterns.
• Acoustic emission variance serves as a reliable fingerprint of fault zone stress, even during aperiodic failure, confirming its physical relevance.
• Pretraining Transformer models on synthetic sine waves improved training stability and performance on limited real lab data, suggesting a viable strategy for data-scarce scenarios.