A data-driven approach to earthquake early warning: Multicomponent site-spectra prediction using deep neural networks

Document Type

Article

Publication Date

3-2026

Abstract

This paper presents a hybrid deep learning framework for earthquake early warning (EEW) that leverages front-site observations to predict target-site spectral characteristics—specifically Fourier amplitude spectra (FAS) and 5% damped pseudo-velocity response spectra (pSᵥ) in real time. In its current form, the framework is site-specific, as the front-site/target-site pairs used for training and evaluation are fixed. By integrating a convolutional neural network (CNN) front end with a long short-term memory (LSTM) sequence model, our approach captures both spatial frequency content and temporal correlations without requiring explicit source, path, or detailed geological inputs. Trained on a diverse corpus of historic accelerograms, the CNN-LSTM network learns cross-spectral and multicomponent dependencies and region-specific site effects, yielding rapid, physically consistent spectral estimates. We evaluate its performance across five case studies, demonstrating that our model not only reduces prediction error relative to established GMPEs for both FAS and pSᵥ, but also preserves spectral shape and cross-period correlations essential for reliable EEW. The developed technique is capable of estimating target-sites through very low latency inference, providing real-time capabilities. Compared to traditional GMPE-based warnings, our data-driven method achieves substantially faster issuance and improved shaking intensity forecasts. We conclude by outlining avenues for embedding sites’ distance and physics-informed constraints, expanding observation datasets, and enhancing model usefulness in seismic demand prediction which are key steps toward rapid EEW systems.

Link to Full Text

Share

COinS