The Accuracy of Deep Learning Versus ETAS in Earthquake Aftershock Forecasting

Emergency responders and civil defense agencies rely on aftershock forecasts to decide when it is safe to enter a collapsed structure or reopen a bridge. For decades, those probabilities have been calculated by the Epidemic-Type Aftershock Sequence (ETAS) model, a statistical workhorse developed in 1988 that assumes earthquakes trigger offspring according to fixed empirical laws. Over the last eight years, seismology has seen a massive influx of deep learning models promising to revolutionize these forecasts by learning complex, hidden triggering patterns directly from massive seismic catalogs.

However, a systematic review of the evidence reveals a surprising plateau. When benchmarked rigorously against a properly calibrated ETAS model, the most advanced neural networks have not yet demonstrated a statistically significant advantage in predictive accuracy. The true breakthrough of artificial intelligence in earthquake forecasting is not in seeing the future more clearly, but in calculating it exponentially faster.[6]

The debate over artificial intelligence's predictive superiority began in August 2018, when a high-profile study published in Nature deployed a deep neural network trained on 131,000 mainshock-aftershock pairs. The model achieved an Area Under the Curve (AUC) of 0.849 in forecasting aftershock locations, significantly outperforming a classic Coulomb failure stress change model, which scored 0.583. It was heralded as a milestone for machine learning in geophysics.[1]

That milestone was quickly contextualized. In October 2019, researchers Arnaud Mignan and Marco Broccardo demonstrated that a simple logistic regression model—effectively a neural network with a single neuron—could achieve the exact same 0.849 AUC when fed the same spatial data. The deep learning model had successfully identified the distance from the mainshock rupture as the primary driver of aftershocks, but its 13,000 parameters provided no additional predictive power over a two-parameter statistical baseline.[2]

Since then, the architectures have grown vastly more sophisticated. Researchers have moved beyond static spatial maps to Neural Point Processes (NPPs), which model the continuous time evolution of an earthquake sequence. In August 2023, the Recurrent Earthquake foreCAST (RECAST) model demonstrated that deep learning could successfully ingest large, diverse earthquake catalogs and independently recover the statistical patterns expected in seismicity, offering a highly flexible framework.[3]

Yet, when subjected to strict, pseudo-prospective testing—where models are evaluated on data strictly subsequent to their training period—the accuracy gap remains stubbornly closed. The March 2026 EarthquakeNPP benchmark, which standardized evaluation protocols across the seismology and machine learning communities, tested five modern neural point processes against ETAS. Using precision-recall AUC (PR-AUC), the ETAS baseline scored 0.125, outperforming the best neural classifier, which scored 0.101. The authors concluded that "none of the five NPPs tested outperform ETAS," suggesting that current implementations are not yet ready to replace the 1988 standard.[5]

Where deep learning definitively wins, however, is in operational latency. The ETAS model requires running thousands of computationally expensive simulations to generate a probabilistic forecast, a process that can take hours or even days for a large, active sequence on standard hardware. In contrast, a trained neural network can output a spatial probability map almost instantly. As Foteini Dervisi, lead author of a November 2025 study on spatiotemporal forecasting, noted, "machine learning models can produce aftershock forecasts within seconds, showing comparable quality to that of ETAS forecasts."[4]

This speed differential fundamentally changes how forecasts can be deployed. In the immediate aftermath of a magnitude 7.0 mainshock, when the risk of a devastating secondary rupture is highest, waiting a day for an ETAS simulation is operationally useless. A neural network that matches ETAS accuracy but delivers the map in four seconds allows emergency managers to make immediate, data-driven decisions about evacuation zones and rescue operations.

The evidence suggests that the next leap in earthquake forecasting will not come from algorithmic complexity alone, but from the data feeding it. Machine learning is currently being used to build high-resolution seismic catalogs that detect millions of micro-earthquakes previously lost in the noise. As these enhanced datasets grow, neural models—unconstrained by the fixed mathematical kernels of ETAS—may finally find the hidden spatiotemporal structures required to break the accuracy ceiling.