Presentation Information
[S08-03]Deep learning-based forecast of final moment magnitude in dynamic earthquake sequence simulation
*Reiju Norisugi1, Yoshihiro Kaneko1, Bertrand Rouet-Leduc2 (1. Graduate School of Science, Kyoto University, 2. Desaster Prevention Research Institute, Kyoto University)
Recent advances in machine learning (ML) and deep learning (DL) have enabled the prediction of shear-slip failures (lab quakes) in centimeter-scale friction experiments and submeter-scale gels. ML or DL, typically trained on continuous acoustic emissions or event catalogs, predicts time-to-failure or applied shear stress. Despite this progress, whether the final size (e.g., moment magnitude) of the next event can be predicted before the rupture is rarely addressed. Even if the timing of the next event is predicted by the ML approach, it is crucial to estimate how large the next event growth is important for the earthquake forecasting.
A few studies using shear experiment of submeter-scale gel material in a tectonic-like setting indicate that the trained models implicitly infer the next event size. They output spatial map of time to failure, and the area where the time to failure is short clearly correlates to the next failure region. However, it has not yet assessed whether the quantitative forecast of the event size is possible, and it is also unclear how the fault constitutive parameters or heterogeneity affect the ability of forecasting. In far-field observation of seismic wave, it is currently understood that the final moment magnitude of natural earthquakes can be speculated when they start to cease. However, it remains elusive whether this limitation reflects an intrinsic property of the physical system itself, or whether predictive information exists that has not yet been fully recognized or explored.
To address the question, we applied an advanced DL technique (convolutional neural network, CNN) to dynamic earthquake sequence simulations with rate-and-state friction and examine whether the final magnitude of an earthquake can be predicted before the rupture onset. We imagine Nankai earthquake sequence and prepare the simple fault models that has a velocity-strengthening barrier at the center and assume variety of characteristic weakening distance of the fault. With relatively strong barrier (large frictional parameter a-b), many dynamic events are arrested near the fault center, and some propagates through the barrier, being large event. Besides, with relatively small weakening distances, the earthquakes can be arrested before it reaches the velocity-strengthening barrier, being small events. Both effects make earthquake sequence complex. At first, to investigate the intrinsic property of earthquake sequence which is frequently used to reproduce laboratory and natural earthquakes, we input the slip rate distribution on the fault to the CNN and train to concurrently output the final potency (a proxy of moment magnitude) and time to the next failure. We find that the CNN can accurately predict the final potency and time to failure in a simple setting but has difficulty in predicting them in such complex sequences, though the parameter distribution is still simple and it is an ideal situation that the fault slip rate can be used. The large events are relatively well predicted in a complex sequence, however, notable misprediction arises when small events occur. We also use quasi-dynamic earthquake sequence simulation, which neglect the wave propagation effects, and find that the problems become less challenging. This result implies that accounting for dynamic effects is crucial to understand the predictability of final moment magnitude. We are currently attempting to understand how the trained CNN infer the final event size by analyzing the model weights. Our results suggest that probabilistic prediction of event size is feasible, but highly sensitive to the complexity of the fault’s physical properties.
A few studies using shear experiment of submeter-scale gel material in a tectonic-like setting indicate that the trained models implicitly infer the next event size. They output spatial map of time to failure, and the area where the time to failure is short clearly correlates to the next failure region. However, it has not yet assessed whether the quantitative forecast of the event size is possible, and it is also unclear how the fault constitutive parameters or heterogeneity affect the ability of forecasting. In far-field observation of seismic wave, it is currently understood that the final moment magnitude of natural earthquakes can be speculated when they start to cease. However, it remains elusive whether this limitation reflects an intrinsic property of the physical system itself, or whether predictive information exists that has not yet been fully recognized or explored.
To address the question, we applied an advanced DL technique (convolutional neural network, CNN) to dynamic earthquake sequence simulations with rate-and-state friction and examine whether the final magnitude of an earthquake can be predicted before the rupture onset. We imagine Nankai earthquake sequence and prepare the simple fault models that has a velocity-strengthening barrier at the center and assume variety of characteristic weakening distance of the fault. With relatively strong barrier (large frictional parameter a-b), many dynamic events are arrested near the fault center, and some propagates through the barrier, being large event. Besides, with relatively small weakening distances, the earthquakes can be arrested before it reaches the velocity-strengthening barrier, being small events. Both effects make earthquake sequence complex. At first, to investigate the intrinsic property of earthquake sequence which is frequently used to reproduce laboratory and natural earthquakes, we input the slip rate distribution on the fault to the CNN and train to concurrently output the final potency (a proxy of moment magnitude) and time to the next failure. We find that the CNN can accurately predict the final potency and time to failure in a simple setting but has difficulty in predicting them in such complex sequences, though the parameter distribution is still simple and it is an ideal situation that the fault slip rate can be used. The large events are relatively well predicted in a complex sequence, however, notable misprediction arises when small events occur. We also use quasi-dynamic earthquake sequence simulation, which neglect the wave propagation effects, and find that the problems become less challenging. This result implies that accounting for dynamic effects is crucial to understand the predictability of final moment magnitude. We are currently attempting to understand how the trained CNN infer the final event size by analyzing the model weights. Our results suggest that probabilistic prediction of event size is feasible, but highly sensitive to the complexity of the fault’s physical properties.