Original title: Adapting a Pretrained LLM to Rapidly Recover Earthquake Magnitude and Location
Aurora Bassani, Daniele Trappolini, Giulio Poggiali, Elisa Tinti +3 · September 2026
When earthquakes strike, automated systems must quickly figure out where they started and how strong they were. Researchers tested whether Large Language Models, which are AI systems usually trained on text, could handle this task. They adapted a small model called TinyLlama using a fine-tuning method named LoRA. To test rapid detection, the model uses data only from monitoring stations where the first seismic wave—the fast-moving P-waveThe primary seismic wave that travels fastest through rock and arrives first at a seismometer.—arrives within 5 s of the first station triggered. It reads station locations, P-wave arrival times, and peak ground velocities from a tiny 0.2 s window. Without using a model of underground rock speeds, it accurately estimates the earthquake magnitude, location, and origin time. Compared to automated systems used by INGV in Italy, it matches their magnitude estimates and delivers better accuracy for origin time and location, proving that language models can serve as effective seismic tools.
You're trying to guess where a flash of lightning struck just by timing the thunder from two or three nearby microphones. Normally you'd need to know exactly how fast sound moves through different parts of the city—through parks, buildings, over water—to calculate the spot. But if the weather and buildings change every time, that map stops helping. Here the researchers skip the map completely: they feed the microphone positions and the exact moment each one hears the thunder into a small pattern-matching system, and it learns to pinpoint the strike from experience alone, without ever needing to know what the sound traveled through.
When an earthquake hits, every second counts for early warnings and emergency response. This paper shows a small AI model can estimate an earthquake's size and location using just two-fifths of a second of data from nearby stations—without needing a geologic velocity model. That could make automated alerts faster and more accurate, especially in regions where traditional methods are slow or imprecise.
The evaluation is limited to Italian earthquakes recorded by INGV, so how well the approach generalizes to other tectonic regions or station networks is not established. The model currently requires station coordinates and P-wave arrival times, meaning it still depends on automated phase-picking. The abstract does not report performance on larger, more distant, or offshore earthquakes, nor does it assess the model's false-alarm rate or ability to handle missing data.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Adapting a Pretrained LLM to Rapidly Recover Earthquake Magnitude and Location — Aurora Bassani, Daniele Trappolini, Giulio Poggiali et al. — Journal of Geophysical Research Machine Learning and Computation — doi:10.1029/2026jh001578
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.