Climate · AI · Disaster Prediction
Neural networks can now detect the tectonic signals that precede megathrust earthquakes. AI can forecast the floods. But what happens when the disaster is one that has never happened before?
In the hours before a megathrust earthquake, the ground moves. Not in a way anyone can feel. Not in a way that sets off any alarm. But the tectonic plate above a subduction zone deforms - subtly, measurably, in patterns that a neural network trained on thousands of past events can learn to recognize. The machines are getting faster at reading the Earth. The harder question is whether they can read disasters that have never happened yet.
Megathrust earthquakes - the largest earthquakes on Earth, produced where one tectonic plate slides beneath another - are responsible for the most destructive seismic events in recorded history. The 2004 Indian Ocean earthquake and tsunami killed over 200,000 people. The 2011 Tohoku earthquake and tsunami, which triggered the Fukushima nuclear disaster, reached magnitude 9.1. The largest earthquake ever recorded, Chile's 1960 Valdivia event, reached magnitude 9.5.
Predicting these events has been one of the hardest problems in geoscience. Earthquakes are the result of stress accumulating along fault planes over years to centuries, then releasing catastrophically in seconds. The accumulation is gradual and largely invisible. The release is sudden and devastating. Most earthquake prediction research has focused on statistical probability - where and when, within broad windows, are the odds of a large event highest - rather than on identifying specific precursors that reliably indicate imminent rupture.
A 2025 study published in Geophysical Research Letters, led by researcher Graciosa and colleagues, has taken a new approach. The team trained a convolutional neural network using data from laboratory models of subduction zones - scaled-down physical simulations of the tectonic environment in which megathrust earthquakes occur. The model was given deformation data from the overriding plate: the subtle, measurable movement of the surface above the fault, recorded at multiple points across its width.
The network's task was to learn to forecast the time-to-failure (TTF) - the time remaining before the next earthquake rupture - from the pattern of deformation in the plate. What the model found was striking: certain patterns of deformation in the overriding plate appear consistently hours to months before rupture. The plate "knows," in some mechanical sense, that it is about to fail. The neural network can read that knowledge.
The researchers are careful about what this does and does not mean. The network cannot yet forecast earthquakes independently in real-world settings. The laboratory models, while physically faithful to the dynamics of real subduction zones, are not identical to the messy, complex crust of the Pacific Ring of Fire. The patterns the model identifies need to be validated against field data from actual monitoring networks before they can serve as an operational early warning signal.
But the proof of concept is real and significant. If the overriding plate deforms in detectable ways hours to months before rupture, then dense monitoring networks at subduction zones - GPS sensors, tiltmeters, strain gauges embedded at strategic intervals along the fault - feeding data in real time to neural networks trained on the laboratory patterns, could provide warning windows that current science cannot. For the populations living in the shadow of the Cascadia Subduction Zone, the Nankai Trough, and the dozen other major megathrust systems around the Pacific, even hours of advance warning translates directly into lives.
The transformation of weather forecasting by artificial intelligence is already well underway. Models like Google's GraphCast and Huawei's Pangu-Weather can generate global forecasts at roughly 45,000 times the speed of traditional numerical weather prediction systems - and in many head-to-head comparisons, they match or surpass the accuracy of those systems. They are trained on decades of reanalysis data: gridded, historical records of atmospheric conditions that give the models a rich statistical picture of how weather behaves.
In March 2026, researchers at the University of Minnesota published paired studies demonstrating that "knowledge-guided" AI - systems that incorporate physical laws and domain expertise alongside data-driven pattern recognition - can meaningfully improve flood forecasting. The models combine the statistical power of machine learning with the physical constraints that pure data approaches sometimes violate, producing predictions that outperform either approach in isolation. For river basins where flood risk is driven by rainfall events, these models offer the prospect of longer warning windows and more accurate predictions of peak flow levels.
A 2025 review in Nature Communications surveyed the state of AI across disaster categories - floods, droughts, wildfires, heatwaves - and found consistent evidence that machine learning approaches outperform traditional statistical models in prediction tasks where enough historical data exists to train them. The review noted both the advances and the limitations, calling for more transparent and reliable AI systems to build the kind of stakeholder trust that operational use requires.
There is a concept in risk analysis called the "black swan" - an event so unprecedented that it falls entirely outside any historical model, a surprise by definition. A related concept, the "gray swan," describes events that are theoretically possible but so rare that they have never appeared in any dataset. Not impossible. Not unimaginable. Simply absent from the training record.
For AI weather models, gray swans represent a fundamental vulnerability. A 2025 study published in Proceedings of the National Academy of Sciences tested this directly. Researchers took an AI weather model and deliberately removed Category 3-5 tropical cyclones from its training data, then asked it to forecast actual Category 5 storms. The model failed. Confronted with an event outside its training distribution - one it had never "seen" in any form - it produced inaccurate forecasts.
This is not a trivial limitation. Climate change is increasing the frequency and intensity of extreme weather events across multiple categories - floods, heatwaves, extreme rainfall, powerful cyclones. Some of these events will exceed any threshold seen in the historical record. A weather model trained on the past will not, by construction, have learned from them. The researchers found some encouraging signs: models trained on strong storms in one region could sometimes transfer that knowledge to forecast them in another region with similar atmospheric dynamics. But the fundamental problem - AI systems struggling with events outside their training distribution - remains unsolved.
The societal stakes are rising as AI weather models move from experimental tools to operational systems. Emergency services, infrastructure managers, and governments are beginning to rely on AI forecasts. If those forecasts systematically underestimate unprecedented events - precisely the events that climate change is making more likely - the consequences of that failure will fall hardest on the communities least able to absorb them.
The limitations of pure data-driven AI in the face of unprecedented events point toward a solution that is becoming a consensus in the field: hybrid models that fuse the pattern-recognition power of machine learning with the explanatory grounding of physical science.
The University of Minnesota's "knowledge-guided" AI for flood forecasting is one example. By encoding physical laws - conservation of mass, the hydraulic equations that govern how water flows through channels - directly into the model architecture, the researchers constrain the predictions to remain physically plausible even when they are extrapolating beyond the training data. A purely statistical model, faced with a rainfall event outside its training distribution, might produce a flood forecast that violates basic principles of hydrology. A physics-informed model cannot make that error, because the physics is built in.
The same principle is driving progress in earthquake science. The neural networks now being trained on subduction zone models are not operating in a data vacuum - they are informed by decades of theoretical understanding of tectonic mechanics. The model learns the patterns; the physics provides the framework within which those patterns make sense. The explainability that results - the ability to say not just "the model predicts imminent rupture" but "it predicts this because the plate is deforming in this specific way, which the physical theory predicts should precede failure" - is essential for the kind of institutional trust these systems need to achieve operational deployment.
Earthquake early warning systems already save lives. Japan's national system, the most sophisticated in the world, can detect an earthquake in progress and transmit warnings that arrive at distant locations seconds before the shaking does - enough time to brace, to stop trains, to halt surgeries. The 2011 Tohoku earthquake's warnings reached some parts of Japan 15 seconds before the strongest shaking arrived. In a hospital, 15 seconds is enough to stop a scalpel mid-surgery. In a factory, it is enough to stop hazardous machinery.
The new AI work on megathrust precursors operates on a different timescale: not seconds, but hours to months. A warning window of that length is transformative. It is enough to evacuate coastal zones. Enough to pre-position emergency resources. Enough to shut down nuclear reactors along the coast, to move hospitals' most vulnerable patients to higher ground, to issue credible public warnings that leave time for deliberate response rather than panic.
For flood prediction, the gains are similarly tangible. The difference between a 6-hour flood warning and a 24-hour flood warning is the difference between people fleeing with their bodies and people evacuating with their possessions, their medications, their documents, their vehicles. It is the difference between livestock losses and livestock moved to safety. For the 1.8 billion people globally exposed to significant flood risk, improvements in prediction lead directly and verifiably to reductions in death, injury, and economic loss.
AI is not going to prevent the next great earthquake or tame the next Category 6 hurricane. But it may be the tool that gives us enough warning to not be standing in its path when it arrives.
Lisa PedrosaThe trajectory of both the problem and the solution is clear. Climate change is loading the atmosphere with more energy, increasing the frequency and intensity of extreme precipitation events, warming ocean surfaces that fuel tropical cyclones, extending fire seasons, and disrupting the jet streams that historically shaped weather patterns in the temperate zones. The number of weather-related disasters recorded each year has roughly tripled over the past fifty years, according to the World Meteorological Organization. The gray swans are getting less rare.
At the same time, the AI tools being built to forecast these events are improving rapidly. The knowledge-guided models being developed for flood prediction are demonstrably better than what preceded them. The neural networks being trained on subduction zone models represent a genuinely new approach to a problem that has resisted solution for decades. The integration of physical understanding with data-driven pattern recognition is producing systems that can extrapolate more reliably than either approach alone.
The challenge is not only technical. It is also institutional - about who uses these tools, how they are validated, what happens when they fail, and how their outputs are communicated to decision-makers and the public in ways that lead to useful action rather than confusion or false reassurance. The failure mode for AI weather prediction is not subtle: a model that systematically underforecasts an unprecedented event, in a world where that model is trusted, produces exactly the wrong outcome at exactly the wrong moment.
The researchers working in this field are aware of this. The calls for transparency, for rigorous out-of-distribution testing, for hybrid approaches that combine the speed of AI with the physical grounding of traditional meteorology, reflect a scientific community that is trying to build tools worthy of the stakes. The ground is already moving, in more ways than one. The hour before the wave may be the most important hour we have.
© 2026 Lisa Pedrosa · lisapedrosa.com
All articles cited to primary institutional or peer-reviewed sources
Buy me a coffee