Neural networks crack hidden friction patterns that drive slow earthquakes

Deep beneath southwest Japan, a fault is quietly slipping. Not with the violent jolt of a great earthquake, but with a slow, silent creep that unfolds over weeks and months — a phenomenon scientists call a slow slip event. For decades, researchers have suspected that these enigmatic tremor-less ruptures and their fast, destructive cousins share a common origin: patchy, uneven friction along the same faults. Now, a team of Japanese and American scientists has borrowed a tool from the artificial intelligence revolution to prove it in unprecedented detail, using physics-informed neural networks to read the frictional fingerprint of a real slow slip event directly from ground measurements.

The study, published in Communications Earth & Environment, was led by Rikuto Fukushima of Stanford University, together with Masayuki Kano of Kyoto University and the Japan Agency for Marine-Earth Science and Technology, Kazuro Hirahara of RIKEN and Kagawa University, and Makiko Ohtani of Kyoto University. Their target was the 2010 Bungo slow slip event, a well-documented episode of silent fault motion that struck the Bungo Channel region between the islands of Shikoku and Kyushu. What the team found was a striking picture of frictional heterogeneity: the slow slip event nucleated in a localized patch off southwest Shikoku, where frictional properties differ sharply from their surroundings, and then propagated westward along the plate boundary.

To understand why this matters, it helps to grasp the long-standing puzzle of fault behavior. Laboratory experiments on rock samples have shown that faults obey friction laws — mathematical rules describing how resistance to sliding depends on slip rate, contact time, and other variables. When modelers feed these laboratory-derived laws into simulations of fault slip, the simulations naturally produce a rich diversity of behavior: some patches rupture rapidly as earthquakes, others creep slowly, and still others remain locked for centuries, accumulating the strain that eventually powers the largest quakes. The agreement between such simulations and real observations has long supported the hypothesis that the observed diversity of slip — from fast earthquakes to slow transient events — arises from spatial variations in frictional properties stamped into the fault by geology.

But there has been a stubborn gap between that elegant hypothesis and hard data. Confirming it quantitatively would require taking real geodetic observations — the millimeter-to-centimeter shifts of GPS stations as an unseen fault slips below — and working backwards to infer the frictional parameters that must have produced them. That inverse problem is brutally difficult. A fault plane is a two-dimensional surface, and frictional properties vary across it point by point, meaning the unknowns number in the thousands or more. Traditional optimization methods buckle under this high-dimensional search space, so most studies have settled for estimating only the distribution of fault slip itself, leaving the underlying friction physics unconstrained. The new work was designed precisely to close that gap.

The team’s solution was to deploy physics-informed neural networks, or PINNs, a technique that has swept through computational physics in recent years. In a conventional neural network application, the network learns patterns purely from data. In a PINN, the network is additionally constrained to obey the governing equations of the physical system — in this case, the equations of quasi-dynamic fault slip based on rate- and state-dependent friction laws. The researchers used the neural network to represent the spatially variable frictional properties on the fault surface. Because a neural network can encode smooth, structured fields with relatively few internal parameters, it tames the high-dimensional optimization problem that had defeated earlier approaches. The network’s outputs are fed into the fault mechanics equations, and the mismatch between the resulting predicted surface displacements and the actual GPS observations drives the training, pulling the estimated friction field toward one that is simultaneously mechanically consistent and consistent with the data.

Applying this framework to the 2010 Bungo slow slip event, the researchers estimated the heterogeneous frictional parameters along the subducting plate interface beneath southwest Japan. The resulting picture showed frictional properties varying systematically along the fault, with a localized region off southwest Shikoku standing out as the nucleation zone where the slow slip event began. From there, the event propagated westward, a pattern that emerged naturally from the estimated friction distribution. Crucially, the calculated fault slip produced by those frictional properties successfully reproduced the spatio-temporal pattern of the observed surface displacements — not just how much the ground moved, but the sequence and geography of that motion through time.

This dual consistency is what sets the result apart. Many previous slip models could fit geodetic data, but they said nothing about whether the inferred slip was physically achievable given plausible fault friction. Conversely, forward simulations of fault cycles could produce realistic slow slip events, but only for assumed friction distributions, not ones verified against observations. The PINN-based inversion delivers a model that is validated on both fronts: it is a mechanically consistent fault slip model, grounded in laboratory-derived friction physics, and it is quantitatively checked against real measurements. In effect, the team has demonstrated that the frictional heterogeneity hypothesis — the idea that patchy friction generates the spectrum of slip behavior — can be tested against data rather than merely assumed.

The study did not stop at reconstructing the past. The researchers also used their estimated friction model to forecast the future evolution of fault slip, and in doing so they uncovered a practical lesson for the emerging field of earthquake forecasting. The accuracy of such forecasts depends heavily on how much observational history is assimilated into the model. By assimilating observations spanning multiple slow slip event cycles, rather than a single event, the forecasts become substantially more reliable. This finding carries a clear message for monitoring networks: continuous, long-term geodetic observation is not just an archive of past behavior but an essential ingredient for anticipating future fault motion.

The implications reach well beyond the Bungo Channel. Slow slip events are now recognized along many of the world’s subduction zones, including Cascadia in the Pacific Northwest, Mexico, Chile, New Zealand, and Alaska. In several of these regions, slow slip episodes have been linked to heightened stress on adjacent locked patches capable of producing large earthquakes — most notably in the lead-up to the 2011 Tohoku-oki earthquake in Japan, where precursory slow slip near the hypocenter has been documented. If physics-informed inversion can map the frictional architecture of these faults from routine GPS data, it could eventually help scientists identify which slow slip events are routine and which signal dangerous stress transfer toward locked, earthquake-ready patches.

The methodological advance may prove just as consequential as the specific findings. High-dimensional inverse problems pervade geophysics, from imaging Earth’s interior to modeling magma chambers and groundwater flow. By showing that a neural network can serve as a compact, physically constrained representation of a complex field like fault friction — and that training it against both physical laws and observational data yields a solution previously out of reach — the study opens a template that other researchers can adapt. The work was supported by Japan’s MEXT STAR-E and STAR-E NEXT projects, the Earthquake Research Institute at the University of Tokyo, and JSPS KAKENHI grants, reflecting a sustained national investment in earthquake science and machine learning.

For a field that has spent half a century bridging laboratory rock mechanics and field observation, the arrival of a tool that finally connects the two at scale is a genuine milestone. The 2010 Bungo slow slip event slipped silently, registering only as a gentle drift of GPS antennas across Shikoku and Kyushu. Yet within that quiet drift lay the frictional blueprint of the fault itself — and now, with physics-informed deep learning, scientists have learned to read it. The same approach, applied to longer and richer observational records, promises not only a deeper understanding of how faults work but a step toward the long-sought goal of physics-based fault slip forecasting, where the slow whispers of a fault today inform the assessment of its roar tomorrow.

Subject of Research: Physics-informed deep learning estimation of spatially variable fault friction from geodetic observations of slow slip events

Article Title: Physics-informed deep learning links geodetic data and fault friction

Article References: Fukushima, R., Kano, M., Hirahara, K., & Ohtani, M. (2026). Physics-informed deep learning links geodetic data and fault friction. Communications Earth & Environment. https://doi.org/10.1038/s43247-026-04138-y

Image Credits: AI Generated

DOI: 10.1038/s43247-026-04138-y

Keywords: physics-informed neural networks, slow slip events, fault friction, geodesy, subduction zone, Bungo Channel, rate-and-state friction, machine learning, seismology, earthquake forecasting, southwest Japan, fault mechanics