Alerting for imminent earthquakes

Forecasting the occurrence of earthquakes, particularly of high magnitude (≥5), is a notoriously difficult problem (Nat. Comput. Sci. 3, 115; 2023). Often, scholars devote their efforts to detect and make use of possible signals that may warn of an impending earthquake — such as surface deformations prior to a shock — but observing and collecting data about these signals can be time-consuming and laborious. Recent developments in geodetic techniques, however, have enabled real-time and millimeter-scale monitoring of the movements of the Earth’s surface, which generates high-dimensional spatial geodetic time-series data in a timely and robust manner that could be used to improve the accuracy of earthquake-warning systems. In a recent work, Pei Chen, Rui Liu, Luonan Chen and colleagues leverage this data and a reservoir computing framework to generate a model-free method that is capable of alerting for possible catastrophic seismic events with better performance than previous methods, suggesting its potential for practical real-time early warning of high-magnitude earthquakes.

In their proposed model, the authors take advantage of the Auto-Reservoir Neural Network (ARNN), an approach that integrates reservoir computing (which is suitable for temporal/sequential information processing) and spatiotemporal information transformation (STI). STI can effectively transform the spatial information of high-dimensional data (such as the geodetic data) to the temporal dynamics of any target variable (such as the target location of the earthquake) and latent variable (such as the center manifold of the geodynamical system near the tipping point of the earthquake). Any inconsistencies in the prediction of the target variable and any strong fluctuation of the latent variable implies an imminent critical transition/earthquake. The authors applied their model to real observations from five different high-risk earthquake regions, demonstrating an overall accuracy of about 98% for detecting earthquakes of magnitude ≥5; perhaps most importantly, the false positive rate was less than 1%, and the early signals were identified 6.27 days prior to the earthquakes on average. The authors also showed that their model outperformed ten other existing methods, and that it worked better in alerting for earthquakes with shallow focal depths, which are often more destructive. Given the challenging task of forecasting earthquakes, the proposed model represents an important advance in the field and has the potential to be used in alerting for strong earthquakes.