17:15 〜 19:15
[SSS10-P34] Earthquake Stress-Drop Estimation in Laboratory Experiment with Machine Learning
キーワード：Stress Drop, Earthquake Rupture Processes, Laboratory Earthquakes, Direct Shear Experiments, Machine Learning
Stress drop is a key parameter that influences earthquake rupture processes and ground-motion characteristics. By quantifying the difference in shear stress before and after a rupture, it offers direct insight into how energy is released during an earthquake. However, stress drop varies significantly from fault to fault and from event to event, underscoring the importance of improved measurement techniques. Previous studies have successfully adopted machine learning models in laboratory earthquake datasets to predict time to failure and shear stress. However, their experimental fault systems were predominantly unstable. Specifically, the system is unstable when the loading stiffness k’ is much lower than the critical stiffness kc’, i.e., k’ ≪ kc’. The single state variable rate-and-state friction model suggests that slow-slip behavior emerges in the special case where k’ ~ kc’, which has significance for slow earthquake detection and earthquake early-warning systems.
In this study, we conducted double direct shear experiments on granite faults and generated laboratory earthquakes across a range of critical stiffness ratios (K = k’/kc’), from near-critical conditions (K ~ 1) to highly unstable (K ≪ 1). These conditions were achieved by varying normal stress, shearing rate, and loading stiffness. Acoustic emission (AE) data were collected continuously at 2.5 MHz using an array of 12 ultrasonic piezoelectric transducers embedded in a steel block, which is in direct contact with one of the granite side blocks. Additionally, shear strain was recorded at 13 equidistant locations at 100 kHz to capture transient local strain perturbations along the 15-cm saw-cut granite fault. At near-critical conditions, we observed consecutive events with highly variable stress-drops. The datasets with diverse event characteristics better emulate natural faults without strictly repeating rupture patterns and offer a broader training set for machine learning models, enhancing robustness in stress drop estimation and potentially improving their ability to generalize across varying fault conditions.
For large earthquakes, well-established techniques—such as finite-fault inversions, back-projection, and other seismological models—can be used to estimate stress drop. However, for small earthquakes (magnitude M < 2), these methods often become infeasible due to sparse data and low signal-to-noise ratios. To date, accurately measuring the in-situ stress drop at small fault ruptures remains a challenge, and thus these techniques are typically evaluated through numerical simulations or laboratory analogs. Leveraging the power of Transformer-based machine learning, multiple acoustic emissions signals from different channels are concatenated into multi-dimensional features and fed into a Transformer model. By treating all these signals as one “multi-feature” sequence (rather than separate inputs), the self-attention mechanism within the Transformer can learn the complex temporal correlations among them. The model can effectively capture how changes in one signal relate to another over time, enabling robust predictions of stress drop even with limited, noisy data.
Our results confirmed that Transformer-based machine learning models can effectively estimate stress drop in laboratory earthquakes using AE data. The diverse dataset produced in our experiments better captures the variability seen in natural faults, making it more promising to apply ML models trained on laboratory data to real-world seismic events.
In this study, we conducted double direct shear experiments on granite faults and generated laboratory earthquakes across a range of critical stiffness ratios (K = k’/kc’), from near-critical conditions (K ~ 1) to highly unstable (K ≪ 1). These conditions were achieved by varying normal stress, shearing rate, and loading stiffness. Acoustic emission (AE) data were collected continuously at 2.5 MHz using an array of 12 ultrasonic piezoelectric transducers embedded in a steel block, which is in direct contact with one of the granite side blocks. Additionally, shear strain was recorded at 13 equidistant locations at 100 kHz to capture transient local strain perturbations along the 15-cm saw-cut granite fault. At near-critical conditions, we observed consecutive events with highly variable stress-drops. The datasets with diverse event characteristics better emulate natural faults without strictly repeating rupture patterns and offer a broader training set for machine learning models, enhancing robustness in stress drop estimation and potentially improving their ability to generalize across varying fault conditions.
For large earthquakes, well-established techniques—such as finite-fault inversions, back-projection, and other seismological models—can be used to estimate stress drop. However, for small earthquakes (magnitude M < 2), these methods often become infeasible due to sparse data and low signal-to-noise ratios. To date, accurately measuring the in-situ stress drop at small fault ruptures remains a challenge, and thus these techniques are typically evaluated through numerical simulations or laboratory analogs. Leveraging the power of Transformer-based machine learning, multiple acoustic emissions signals from different channels are concatenated into multi-dimensional features and fed into a Transformer model. By treating all these signals as one “multi-feature” sequence (rather than separate inputs), the self-attention mechanism within the Transformer can learn the complex temporal correlations among them. The model can effectively capture how changes in one signal relate to another over time, enabling robust predictions of stress drop even with limited, noisy data.
Our results confirmed that Transformer-based machine learning models can effectively estimate stress drop in laboratory earthquakes using AE data. The diverse dataset produced in our experiments better captures the variability seen in natural faults, making it more promising to apply ML models trained on laboratory data to real-world seismic events.