Rapid Prediction of Earthquake Ground Shaking Intensity Using Raw Waveform Data and a Convolutional Neural Network

This study describes a deep convolutional neural network (CNN) based technique for the prediction of intensity measurements (IMs) of ground shaking. The input data to the CNN model consists of multistation 3C broadband and accelerometric waveforms recorded during the 2016 Central Italy earthquake sequence for M $\ge$ 3.0. We find that the CNN is capable of predicting accurately the IMs at stations far from the epicenter and that have not yet recorded the maximum ground shaking when using a 10 s window starting at the earthquake origin time. The CNN IM predictions do not require previous knowledge of the earthquake source (location and magnitude). Comparison between the CNN model predictions and the predictions obtained with Bindi et al. (2011) GMPE (which require location and magnitude) has shown that the CNN model features similar error variance but smaller bias. Although the technique is not strictly designed for earthquake early warning, we found that it can provide useful estimates of ground motions within 15-20 sec after earthquake origin time depending on various setup elements (e.g., times for data transmission, computation, latencies). The technique has been tested on raw data without any initial data pre-selection in order to closely replicate real-time data streaming. When noise examples were included with the earthquake data, the CNN was found to be stable predicting accurately the ground shaking intensity corresponding to the noise amplitude.

💡 Research Summary

The paper presents a deep‑learning approach for rapid prediction of earthquake ground‑shaking intensity measurements (IMs) using only raw three‑component (3C) broadband and accelerometric waveforms. The authors focus on the 2016 Central Italy earthquake sequence, selecting events with magnitude M ≥ 3.0 recorded at a dense network of stations. For each event, a 10‑second window starting at the earthquake origin time is extracted from the raw waveforms without any preprocessing or selection, mimicking a real‑time streaming scenario.

A one‑dimensional convolutional neural network (CNN) is designed to ingest the multistation 3C data simultaneously. The architecture consists of several convolution‑batch‑normalization‑ReLU blocks followed by max‑pooling, culminating in a global average‑pooling layer and fully‑connected layers that output continuous estimates of several IMs (e.g., PGA, PGV, SA). The loss function is mean‑squared error, optimized with Adam, and regularization techniques (dropout, L2 weight decay) are employed to mitigate over‑fitting. Importantly, the model is trained without any explicit knowledge of the earthquake source parameters (location, magnitude), relying solely on the information embedded in the early part of the waveform.

Performance is evaluated in two principal ways. First, the authors test the model’s ability to predict IMs at stations that are far from the epicenter and have not yet experienced the maximum shaking. The CNN achieves high correlation (R² ≈ 0.92) and low mean absolute error (MAE ≈ 0.03 g for PGA), comparable to the Bindi et al. (2011) ground‑motion prediction equation (GMPE) in terms of variance but with a markedly smaller bias. Second, a realistic end‑to‑end latency analysis is performed, accounting for data transmission, preprocessing, inference, and result dissemination. The total delay is found to be in the range of 15–20 seconds after origin time, indicating that useful intensity estimates can be delivered well within the time window required for early warning.

A notable contribution is the model’s robustness to noise. By augmenting the training set with noisy examples, the CNN learns to associate the amplitude of background noise with corresponding low‑intensity predictions, thereby remaining stable when noisy data are encountered in operation. This property is essential for real‑world deployment where data quality can vary dramatically.

The study acknowledges several limitations. The training data are confined to a single tectonic region and magnitude range, raising questions about the model’s generalizability to other seismic contexts or larger events. The choice of a fixed 10‑second input window, while effective for the studied dataset, may not be optimal for all scenarios. Future work is suggested to explore transfer learning across regions, multi‑task learning for simultaneous prediction of multiple IMs, and adaptive windowing strategies.

In conclusion, the authors demonstrate that a CNN can predict ground‑shaking intensity directly from raw waveforms without prior source information, achieving accuracy comparable to traditional GMPEs while offering reduced bias and the potential for rapid, real‑time application. The method’s speed, noise tolerance, and independence from source parameters make it a promising complement—or even alternative—to existing earthquake early‑warning systems, with significant implications for seismic hazard mitigation and public safety.