Can AI accurately predict earthquakes 72 hours in advance from seismic and atmospheric data ?

Could advances in artificial intelligence, trained on seismic and atmospheric data, reliably predict earthquakes up to three days before they occur? The stakes are enormous—timely warnings could transform disaster preparedness worldwide. Yet, what does the science actually say about this possibility?

Background

Earthquake prediction remains one of the most challenging problems in geoscience. Traditional methods rely on statistical analysis of historical seismicity, geodetic measurements of crustal deformation, and precursor signals such as foreshocks, but none have consistently provided reliable short-term forecasts (e.g., days to weeks) ahead of major events (Jordan et al., 2011; Geller et al., 1997; Lomnitz, 1994).

In recent years, machine learning (ML) approaches have been explored to detect subtle, non-linear patterns in seismic data that may precede earthquakes. Studies have used large-scale datasets from dense seismic networks to train deep neural networks capable of identifying anomalies in waveform features, such as temporal clustering, spectral content, or b-value changes (DeVries et al., 2018; Mignan et al., 2021). Some models report improved performance in forecasting aftershock sequences or detecting early-warning signals on regional scales (e.g., Perol et al., 2018; Zhang et al., 2021). However, the physical interpretability of these anomalies remains debated, and rigorous, prospective validations across diverse tectonic settings are limited (van der Elst et al., 2021).

The inclusion of atmospheric data—such as ionospheric disturbances (e.g., total electron content anomalies), radon emissions, or thermal infrared anomalies—has been suggested as potential precursory indicators, drawing from anecdotal and case-study observations (e.g., Pulinets & Ouzounov, 2011). Satellite-based monitoring (e.g., GOES, Swarm) has enabled broader spatial coverage of such signals, and some ML models have attempted to fuse seismic and atmospheric inputs to enhance predictive skill (e.g., Akhoondzadeh & Di Mauro, 2022). Yet, the mechanisms linking atmospheric changes to tectonic stress remain speculative, and robust evidence of causal pathways is lacking (Thomas et al., 2017; Dautermann et al., 2007).

Despite anecdotal reports and isolated case analyses, the broader geophysical community maintains that no validated method exists for predicting the time, location, and magnitude of earthquakes with sufficient accuracy to warrant public warnings (e.g., Nature editorial, 2018). The USGS explicitly states that reliable short-term prediction is not feasible with current understanding and technology (USGS, 2023). While AI may improve detection of subtle patterns, skepticism persists regarding whether these represent true precursors or spurious correlations (e.g., Mignan, 2016). Thus, the frontier lies in distinguishing signal from noise—and in ensuring that any putative predictive signal can be prospectively validated under blind conditions across multiple seismic regimes.

Short-term earthquake prediction—defined as foretelling a specific event hours to days ahead—remains one of seismology’s most challenging goals. Since the 1970s, researchers have probed relationships between geophysical and atmospheric signals (e.g., electromagnetic anomalies, radon emissions, or ionospheric disturbances) and impending tremors, but large, prospectively validated datasets that cover the full 72-hour horizon are scarce. Statistical studies that claim skill at this timescale often do not survive rigorous, out-of-sample testing or have not been replicated across multiple tectonic settings. Deep-learning models that ingest continuous seismic and meteorological streams have shown promise on retrospective datasets—sometimes reporting apparent gains in short-term forecasting metrics—but these advances have yet to translate into operational systems endorsed by major geological surveys. The absence of a universally accepted physical mechanism linking atmospheric signals to rupture nucleation continues to limit the development of reliable, generalizable predictors at the three-day horizon.

— Enriched May 15, 2026