Original title: Is Seismic Forecasting Possible with Physics-based AI?
Victoria Keane, Manolis Veveakis, Thomas Poulet · August 2026
Predicting earthquakes is hard because data noise and normal filtering hide the signals that matter. Deep underground, a sinking tectonic plate loses water. This triggers slow slip eventsSlow, silent movements along a fault that release stress over days to months instead of seconds., which are quiet, repeating slips along a fault that shift the Earth's surface. These movements leave clues before a big quake happens. Researchers built a physics-based attractorIn physics, a set of states a system naturally evolves toward over time., a tool that models the physical rules of subduction and uses artificial intelligence to find hidden patterns in the surface movement data. Testing this in New Zealand's Hikurangi trench, they predicted an earthquake one week early. They found that predictability limits can reach 5-6 weeks into the future with decadal repeatability, showing that these physical processes follow predictable rules.
You are in a noisy kitchen trying to tell if a pot is about to boil over. The steam from the pot mixes with steam from the kettle and the dishwasher, so you cannot see the real signal. In this paper, the ground's slow creep is the steam, the earthquake is the boil-over, and the AI is a filter that understands the physics of cooking and pulls the true steam out of the background noise.
Earthquakes kill when they arrive without warning. If physics-based AI can flag a quake a week ahead, communities in subduction zones get time to prepare, evacuate, and protect infrastructure. The result suggests that prediction may be possible, and it points to a deeper truth: the process may be more deterministic than we assumed.
The abstract reports a single successful prediction in one subduction zone. It does not demonstrate that the method generalizes to other faults or earthquakes, nor does it address false alarms or the practical reliability of a warning system. The claim of determinism is inferred from repeatability, not proven.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Is Seismic Forecasting Possible with Physics-based AI? — Victoria Keane, Manolis Veveakis, Thomas Poulet
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.