Earthquake sequence simulations skip each rupture: a Fourier neural operator learns the pre- to post-event jump, cutting a 20,000-year benchmark from 84.0 to 11.1 hours
Synopsis
A Caltech team hands the costliest part of earthquake sequence simulation, the dynamic rupture phase, to a Fourier neural operator (FNO), a network that learns maps between functions in the frequency domain, so it predicts the post-event fault state directly from the pre-event state while a conventional quasi-dynamic solver advances the slow interseismic periods. On benchmarks of 20,000 years for a homogeneous fault and 40,000 years for a heterogeneous one, the hybrid simulation reproduces the magnitude and recurrence-time distributions of the fully dynamic reference and stays closer to it than the commonly used quasi-dynamic approximation; total compute for the homogeneous case falls from 84.0 to 11.1 hours.
Interpretation
Long earthquake sequence simulations can now use a conventional solver only during the slow interseismic periods and replace each fast rupture wholesale with a learned state jump, completing simulations spanning tens of thousands of years while retaining inertial, wave-mediated stress effects. The prior practice ignored inertial wave-mediated stress transfer across the whole sequence and substituted a quasi-dynamic approximation, because tracking those effects demanded both extra computation and massive storage; the quasi-dynamic approximation can significantly modify sequence simulations. In simulations spanning 20,000 years for the homogeneous fault and 40,000 years for the heterogeneous one, the KL divergence of the magnitude distribution from the fully dynamic reference is 0.1517 (homogeneous) and 0.0381 (heterogeneous) for the hybrid model, versus 0.6468 and 0.1091 for the quasi-dynamic approximation; the hybrid model is likewise closer on recurrence-time distributions.
The hybrid simulator crosses an earthquake event as a unit: the quasi-dynamic solver advances the interseismic period until the maximum slip rate on the fault crosses a threshold that marks event onset, and the trained network then supplies the fault state at event end. The jump map is learnable because inertial wave-mediated effects are negligible before and after an event, so the map does not depend on earlier history and depends only on the slip rate and state variable at event onset. Training data comprise 20,000 pre- and post-event state pairs for each of the homogeneous and heterogeneous faults, split 14,000 for training, 2,000 for validation and 4,000 for testing; on the test set the version with redundant variables has a median relative error of 0.00436 (homogeneous) and 0.00782 (heterogeneous) for logarithmic slip rate.
Adding the shear stress and the change in the stress-transfer functional to the network's inputs and outputs makes the same jump map more accurate and lets the hybrid simulation continue stably. The version using only the two basic quantities, slip rate and state variable, also learns the jump map, but with larger errors. With redundancy, the median relative error for logarithmic slip rate drops from 0.0172 to 0.0091 (homogeneous) and from 0.0128 to 0.0078 (heterogeneous); the median state-variable error drops from 0.0393 to 0.0207 (homogeneous) and from 0.0300 to 0.0173 (heterogeneous). The source's running text and Table 2 give different homogeneous figures, and the reductions quoted here are the ones stated in the running text.
Perspective
The demonstrated scope is a one-dimensional rate-and-state frictional interface under antiplane shear in homogeneous and heterogeneous fault configurations, with statistical comparisons over 20,000-year homogeneous and 40,000-year heterogeneous simulations. The authors propose extensions to three-dimensional faults, models whose interseismic periods include viscoplastic effects, and more complex dynamic weakening mechanisms such as flash heating and thermal pressurization of pore fluids.
The jump map's usefulness rests on inertial wave-mediated effects being negligible before and after events and on the map being approximately Markovian. In the spatial-discretization test, errors rise noticeably for the homogeneous case as the grid is coarsened from 8192 to 512 points, while the heterogeneous case was tested only down to 1024 points. For large-event recurrence times, the KL divergence is 0.2439 for the hybrid model in the homogeneous case, versus 0.1910 for the quasi-dynamic approximation. The source's running text and Table 2 give inconsistent homogeneous median errors before and after adding redundancy, so one of the two must be taken as authoritative.
