Prediction, hazard evaluation, and response to disasters remain severely problematic due to the nonlinear and multiscale nature of crustal behaviour on Earth and the relative sparsity, noise, and heterogeneity of observations. Even with significant improvements in seismology, conventional statistical and physical models still struggle to make short-term predictions, consistently identify precursors, and provide dynamic situational awareness of the state and post-seismic events. In turn, the rapid development of machine learning (ML), deep learning (DL), and large language models (LLMs) has created new opportunities to extract meaningful patterns from diverse datasets, integrate multimodal information, and enable real-time decision-making in earthquake-prone regions. The paper provides an overview of recent advances in AI-based earthquake studies, including environmental precursors, spatiotemporal seismic prediction, ground-motion prediction, multimodal structural damage, and LLM-based knowledge integration. We discuss developments in hydrochemical anomaly detection using ML models developed in the context of long-term hot spring monitoring and highlight improvements in anomaly detection, as well as the challenges posed by varying indicators and time-dependent instabilities. At the world scale, we consider deep architectures that use spherical convolutions and attention to model seismicity on the curved surface of the Earth, showing significant improvements in accuracy, recall, and long-term dependency modeling. Simultaneously, ensemble ML models for peak ground acceleration prediction and SARIMAX-based time-series models with exogenous variables demonstrate how data-driven models can supersede traditional attenuation relationships and capture some fundamental temporal behaviour of seismic processes. Beyond prediction, we consider the growing importance of LLMs as integrative reasoning systems that can combine heterogeneous streams of information, such as textual reports, sensor logs, social media content, and visual signals. These paradigms support the new pipelines of building earthquake emergency knowledge graphs, performing retrieval-based logistics prediction, creating engineering-grade structural damage estimates, and providing real-time situational awareness based on citizen communication. Their increased utility, however, also creates new domain-grounding, bias, interpretability, and reliability issues in high-stakes settings. In these various uses, there are a few common barriers, such as limited model generalization to tectonic settings, insufficient high-magnitude events for training, physical constraints, and uncertainty quantification, all of which can be addressed. These results highlight that future systems are likely best built by blending physical knowledge with data-driven systems, using multimodal sources including seismic, environmental, satellite, geodetic, and social data, and using LLMs as embodiments of agents operating on transparent tools rather than opaque creators. At the end of the paper, the main directions for future research have been identified, including the need for standardized multimodal benchmarks, hybrid physics-ML designs, simulation-based training controls, robust uncertainty estimation methods, and governance systems that are transparent, fair, and reliable. These advances, combined, will no doubt lead to a new generation of AI-modified seismic forecasting and disaster-response structures that are scientifically defensible and operationally feasible, eventually making societies less susceptible to earthquake hazards.
groups
Nahla B. Abdel-Hamid
mail
-
El-Sayed M. El-Kenawy
mail
-
Mohamed M. Abdelsalam
mail
link
https://doi.org/10.54216/MOR.050101