Answers are grounded in this paper and cite their sources.
Earthquakes cause severe casualties and economic losses worldwide, yet deterministic prediction of individual seismic events remains unfeasible. Consequently, modern seismology emphasizes probabilistic forecasting to quantify seismic hazard and uncertainty. While transformer architectures excel in complex time series modeling, their systematic application to natural tectonic earthquake magnitude forecasting remains limited. Here, we develop an enhanced Temporal Fusion Transformer (TFT) framework optimized via a quantile loss function to generate calibrated probabilistic forecasts of the next earthquake magnitude in a global catalog using historical seismicity, engineered temporal features, and spatial clustering. We also evaluate whether augmenting a small seismic catalog using a physically constrained Tabular Variational Autoencoder (TVAE) improves generalization on chronologically future events. Evaluated on held out real earthquakes using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and average quantile loss, our TFT model achieves a typical prediction error statistically comparable to the strongest recurrent attention baseline and significantly lower than a graph based baseline, while uniquely providing well calibrated 80% prediction intervals covering 85.0% of future events. Furthermore, our causal ablation demonstrates that synthetic TVAE augmentation does not improve forecast accuracy on unseen real events, establishing important practical boundaries for generative resampling in statistical seismology.
No takes yet. Share an insight, caveat, or question.
Lodhi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: