Scalable intermediate-term earthquake forecasting with multimodal fusion neural networks
Yumeng Hu, Qi Zhang, Hengshu Zhu, Baoshan Wang, Hui Xiong, Haitao Wang
Abstract
Seismology is witnessing rapid growth in both the volume and variety of earthquake observational data, but current tools for effectively integrating these heterogeneous data remain limited. Here, we propose SafeNet, a scalable deep learning framework designed to address these challenges through the use of multimodal fusion neural networks. SafeNet integrates 282-dimensional seismic indicators from earthquake catalogs, capturing long-, medium-, and short-term seismic patterns, and associates seismic activity with geological information using integrated maps. Its specialized fusion modules and adaptive attention mechanism enable dynamic spatiotemporal information exchange across regions. To validate SafeNet's performance, we conducted a pseudo-prospective test using a 50-year earthquake catalog from China, demonstrating its superior forecasting performance over 13 state-of-the-art models. Additionally, the successful transfer of models trained on the China dataset to the Contiguous and Western United States further highlights SafeNet's scalability.