Spatio-Temporal Earthquake Forecasting with Informed Point-process Optimization (STEF-IPO)

Yizhou Zhang, Zijun Cui, Leila Mizrahi, Yan Liu, & Yehuda Ben-Zion

Published September 8, 2024, SCEC Contribution #13846, 2024 SCEC Annual Meeting Poster #207

Earthquake prediction is a fundamental unsolved problem with significant societal relevance. In recent years, promising research has been conducted on earthquake forecasting, particularly with the development of deep learning models based on temporal point processes. However, the progress relative to the classical Epidemic Type Aftershock Sequence (ETAS) model has been limited. Also, classical formulations of point processes for earthquake sequences typically require integrating the intensity function, which may be hard to compute with an advanced deep learning model using a non-parametric neural network. This introduces some complexity in both model training and inference.

Here, we introduce the Spatio-Temporal Earthquake Forecasting with Informed Point-process Optimization (STEF-IPO) model, following algorithms such as the Attentive Mixture Density Networks developed in the context of social systems [Sharma et al., 2001]. The STEF-IPO model leverages an advanced Transformer architecture with an attention mechanism to capture correlations between earthquakes. It models event likelihoods using a Gaussian Mixture Model (GMM), which inherently incorporates the knowledge of earthquake clustering in space and time. Additionally, the GMM's structure eliminates the need for integral operations, enabling efficient computation. Our initial results demonstrate that STEF-IPO significantly outperforms commonly applied models, such as ETAS and other neural network-based models (e.g., NSTPP [Chen et al., 2020]). In experiments conducted on an observed earthquake catalog for the western US [Trugman and Ben-Zion, 2023 ], STEF-IPO exhibits advanced performance. In the evaluation of event temporal likelihood, STEF-IPO achieves 22.09, outperforming the results of ETAS and NSTPP which are -7.46 and 15.66 respectively. Simultaneously, for spatial likelihood, STEF-IPO achieves 0.7457, outperforming the -8.28604 and -1.4097 of ETAS and NSTPP. In addition, the STEF-IPO model has significant efficiency improvement over NSTPP by 60 times faster on the same testing and validation sets. Updated results will be presented in the meeting.

Key Words
Deep Learning, Neural Temporal Point Process, Earthquake Forecasting

Citation
Zhang, Y., Cui, Z., Mizrahi, L., Liu, Y., & Ben-Zion, Y. (2024, 09). Spatio-Temporal Earthquake Forecasting with Informed Point-process Optimization (STEF-IPO). Poster Presentation at 2024 SCEC Annual Meeting.

Related Projects & Working Groups
Earthquake Forecasting and Predictability (EFP)