Machine-learned acceleration of earthquake sequence simulations
This paper proposes a hybrid modeling framework that combines a conventional quasi-dynamic solver with a trained Fourier neural operator to efficiently simulate earthquake sequences, achieving computational acceleration while accurately reproducing key long-term statistics and wave-mediated inertial effects that standard approximations miss.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer
Earthquakes are not merely sudden, violent shakings of the ground; they are the climax of a slow, decades-long buildup of stress along the cracks in the Earth's crust. Between these violent ruptures, the fault moves imperceptibly slowly, a quiet creep that accumulates energy over years or centuries. When the stress finally overcomes the friction holding the rock together, the fault slips in a fraction of a second, sending out waves of energy that shake the surface. To understand how these cycles work, scientists run computer simulations that track the fault's behavior over thousands of years. However, these simulations face a massive hurdle: the computer must calculate the slow, quiet periods with large steps, but the moment the earthquake starts, it must switch to tiny, frantic steps to capture the rapid waves of energy. This switch makes simulating a full earthquake cycle incredibly slow and expensive, often requiring the computer to remember every single movement from the past to calculate the present, a task that quickly overwhelms even the most powerful machines.
A team of researchers at the California Institute of Technology has found a way to bypass this computational bottleneck without losing the physics that matter. They developed a new method that combines a traditional, slow-moving simulation with a smart, machine-learning shortcut. Instead of forcing the computer to calculate every tiny vibration during the earthquake itself, they trained a specialized neural network to act as a "jump map." This network learned from thousands of short, high-fidelity simulations what happens to a fault immediately before and immediately after a rupture. Once trained, the network can instantly predict the final state of the fault after an earthquake, skipping the need to simulate the chaotic, wave-filled seconds in between. The computer then uses a standard, efficient solver to handle the long, slow periods of stress buildup, only calling upon the neural network when a rupture is detected.
The researchers tested this hybrid approach on two different types of fault models: one with a uniform, smooth section that slips easily, and another with a more complex, patchy section where some parts resist motion more than others. In both cases, the new method successfully reproduced the long-term statistics of fully detailed simulations. It accurately predicted how often earthquakes of different sizes would occur and how much time would pass between them. Crucially, the hybrid model captured the complex, wave-mediated effects that simpler, older methods miss. Traditional approximations often ignore the inertial waves that travel through the rock during a quake, assuming the stress changes happen instantly and locally. The new method, however, retains the influence of these waves because it learned them from the detailed data. As a result, the hybrid model provided a much more accurate picture of earthquake behavior than the standard approximations used today, while running roughly eight to twelve times faster.
The speedup comes from a dramatic reduction in the time spent on the most difficult part of the calculation. In the fully detailed simulations, calculating a single earthquake event took the computer nearly two minutes for the simpler fault and over four minutes for the complex one. In the hybrid model, the neural network performed the same task in a fraction of a second. Over a simulation spanning 20,000 years, this efficiency meant the hybrid model finished in about 11 hours, compared to 84 hours for the fully detailed version. For the longer, 40,000-year simulation of the complex fault, the time dropped from nearly six days to just over 11 hours. Beyond speed, the new method also eliminated the need to store massive amounts of historical data about the fault's past movements, freeing up significant computer memory.
The study confirms that machine learning can be used to accelerate complex physical simulations without sacrificing the accuracy needed to understand real-world phenomena. By learning the "jump" from the start of a rupture to its end, the neural network captures the essential physics of the earthquake while ignoring the computationally expensive details of the wave propagation. This approach allows scientists to run long-term simulations that were previously too costly to attempt, opening the door to better statistical predictions of earthquake sequences. The researchers note that this technique could be extended to three-dimensional faults and more complex physical processes, such as the heating of rock during a slip or the behavior of fluids within the fault. The work demonstrates that by combining the reliability of traditional physics with the speed of learned patterns, it is possible to simulate the Earth's most violent events with a new level of efficiency and insight.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.