Earthquake Prediction from Paleoseismology: a proof of concept based on CNN analysis of data from scaled sismotectonic models

The advent of artificial intelligence (AI) has opened new avenues in geosciences, particularly for earthquake prediction. Deep learning models, especially Convolutional Neural Networks (CNNs), offer promising capabilities to analyze complex data and detect subtle patterns indicative of seismic activity. However, geophysical records span a time interval that is shorter that the duration of large eartquakes cycle, creating a major challenge for training these models.

In this study, we use paleoseismological data, which include multiple seimic cycles (Cascadia and Sumatra zones). Paleoseismological data are often represented as barcodes, where each "bar" represents an earthquake in time and space. We reproduce these barcodes using a scaled seismotectonic model mimicking subduction megathrust earthquake cycles. The simulated sequences include both partial and complete ruptures, representing earthquakes of varying magnitudes. A CNN model is then trained with these barcodes to predict the timing, location along the margin, and magnitude of the next earthquake.

Our results show that the CNN model can reconstruct the complex temporal loading history and accurately predict the timing of future earthquakes. This approach overcomes the limitations of conventional methods based on slip deficit and highlights the potential of paleoseismological data to enhance seismic forecasting strategies.

This work demonstrates the application of deep learning techniques to paleoseismological data as a tool for earthquake prediction. It opens promising perspectives for seismic hazard assessment and the understanding of fault cycles in subduction zones.