A team of researchers from the Hong Kong University of Science and Technology and the Sydney University of Technology has unveiled POSEIDON, a breakthrough artificial intelligence system that successfully combines deep learning and established seismology principles to predict earthquakes and their cascading effects with unprecedented accuracy.
The research, led by Boris Kriuk and Fedor Kriuk, introduces both a new physics-based neural network and the Poseidon dataset, the largest open-source global earthquake catalog to date, containing 2.8 million seismic events over 30 years. This dataset is now open to researchers around the world.
Unlike traditional “black box” machine learning approaches that ignore the fundamental physics that govern earthquake behavior, POSEIDON embeds core seismic laws directly into its architecture. The system incorporates Gutenberg and Richter's magnitude-frequency relationship and Omori and Utsu's aftershock attenuation law as learnable constraints to ensure predictions match established scientific understanding.
“We have demonstrated that respecting physics increases accuracy, not compromises it,” said Boris Kruk of the Hong Kong University of Science and Technology. “POSEIDON delivers state-of-the-art performance across all prediction tasks while producing scientifically interpretable parameters that fall within established seismic ranges.”
The system simultaneously addresses three interrelated challenges that have previously been tackled separately: identifying aftershock sequences, assessing the likelihood of tsunamis, and detecting foreshocks that may precede larger earthquakes. Such a unified approach takes advantage of the unique relationships between these phenomena.
POSEIDON achieved exceptional results in extensive testing, including an AUC score of 0.971 for tsunami detection, even though these events were only 1.14% of the dataset. This was a particularly difficult prediction task due to the extreme class imbalance. The system outperformed gradient boosting, random forests, and traditional neural network baselines across all evaluation metrics.
Fedor Kriuk from the University of Technology Sydney emphasized the practical implications. “By achieving both the accuracy required for operational early warning systems and the transparency required for scientific trust, we are bridging the gap between pure machine learning and physics-based approaches in high-stakes geophysical applications.”
The learned physical parameters provide additional validation of the approach. The Gutenberg-Richter b value of the model converged to 0.752, and the Omori-Utsu parameters reached p = 0.835 and c = 0.1948 days, all within the range established in the seismological literature. Convergence to physically meaningful values occurred naturally during training without sacrificing prediction performance.
The Poseidon dataset includes precomputed energy features, spatial grid indices for efficient geospatial analysis, and standardized quality metrics. Events span the full magnitude spectrum from 0.0 to 9.1, with full geographic coverage across all latitudes and longitudes.
The research team says future studies will consider integrating real-time seismic data, extending continuous probabilistic hazard prediction, and incorporating crustal stress transfer physics. The published dataset is expected to accelerate progress in physics-based earthquake research around the world.
The findings represent a major advance in earthquake prediction, an area that has long struggled to balance the pattern recognition capabilities of modern AI with the physical constraints that govern earthquake behavior. By demonstrating that these approaches work in harmony rather than in conflict, POSEIDON opens new possibilities for reliable, scientifically-based seismic risk assessment systems.
