WaveCastNet: Rapid Wavefield Forecasting for Earthquake Early Warning via Deep Sequence to Sequence Learning

TL;DR

WaveCastNet reframes earthquake ground-motion forecasting as spatiotemporal sequence-to-sequence prediction and introduces ConvLEM, a convolutional long expressive memory backbone, to capture multiscale spatial and temporal dynamics. The model supports both dense wavefield inputs and sparse sensor data, enabling fast, end-to-end forecasts of full waveforms without explicit source parameter estimation, and provides ensemble-based uncertainty estimates for operational decision-making. It demonstrates strong performance on synthetic point-source and finite-fault scenarios and shows zero-shot generalization to real-world data, with subsecond inference and robustness to noise and latency. Limitations include reduced accuracy for large magnitude, finite-fault earthquakes when trained only on point sources and the domain gap between synthetic and real data, motivating future work on broader finite-fault training, domain adaptation, and inference-velocity optimizations for real-time EEW deployment.

Abstract

We propose a new deep learning model, WaveCastNet, to forecast high-dimensional wavefields. WaveCastNet integrates a convolutional long expressive memory architecture into a sequence-to-sequence forecasting framework, enabling it to model long-term dependencies and multiscale patterns in both space and time. By sharing weights across spatial and temporal dimensions, WaveCastNet requires significantly fewer parameters than more resource-intensive models such as transformers, resulting in faster inference times. Crucially, WaveCastNet also generalizes better than transformers to rare and critical seismic scenarios, such as high-magnitude earthquakes. Here, we show the ability of the model to predict the intensity and timing of destructive ground motions in real time, using simulated data from the San Francisco Bay Area. Furthermore, we demonstrate its zero-shot capabilities by evaluating WaveCastNet on real earthquake data. Our approach does not require estimating earthquake magnitudes and epicenters, steps that are prone to error in conventional methods, nor does it rely on empirical ground-motion models, which often fail to capture strongly heterogeneous wave propagation effects.

Figures (23)

• Figure 1: Illustration of the problem setup and our proposed WaveCastNet architecture. (a) Simulation domain in the San Francisco Bay Area. The area of interest is indicated by the black rectangular box. Point-source earthquakes are placed along the thick white line. Known faults are shown in black USGSFault. Black triangles mark the sensor locations used for training under a sparse sensor configuration (see Section \ref{['sec:WFN']}), red triangles denote the ShakeAlert stations, and blue triangles represent additional sensors used for real-data experiments on the 2018 Berkeley earthquake. (b) Schematic of a $6.0$-magnitude earthquake's rupture plane (blue line). Red stars indicate epicenters of two hypothetical earthquakes (A, B) and the 2018 Berkeley earthquake (C). Large red triangles highlight the three sensor locations referenced in the discussion. (c) Example snapshot of the horizontal $X$-component velocity wavefield at $T=16.67$ seconds from a point-source simulation. (d) Schematic of the WaveCastNet forecasting framework, consisting of encoder and decoder modules built from stacked recurrent cells. The embedding layer supports both (i) dense wavefield inputs and (ii) sparse sensor inputs, trained using a random masking strategy (gray pixels) to enable high-resolution forecasting.
• Figure 2: Point-source earthquake prediction. (a) Snapshots of Y-component velocity wavefields at $T = 7.4$, 11.75, 16.1, and 20.45 seconds for (top) ground truth, (middle) predictions from densely and regularly sampled input, and (bottom) predictions from sparsely and irregularly sampled input (triangles denote sensor locations). (b) Map of predicted PGV values and (c) corresponding PGV prediction errors. (d) Map of predicted $T_{PGV}$ values and (e) corresponding $T_{PGV}$ prediction errors. Errors are computed as the difference between predicted and ground-truth values; positive errors indicate overestimation, and negative errors indicate underestimation.
• Figure 3: Waveforms from (a) San Francisco (NC.J020) and (b) San Jose (NP.1788) for a point-source earthquake. The blue lines indicate the ground truth, the red lines show the mean of the predicted waveforms, and the red shaded areas represent three times the standard deviation bands. The gray shaded areas indicate the input time window of 5.7 seconds.
• Figure 4: Uncertainty estimates for dense point-source ground-motion prediction. (a–c) Mean, error, and standard deviation of $\ln{\mathrm{PGV}}$, respectively. (d–f) Corresponding maps for $T_{PGV}$. The error maps (b, e) show the difference between predictions and ground truth. The hole in (f), centered at $X=40$, $Y=38$ km, corresponds to a location where $T_{PGV}$ falls within the input time window and is therefore excluded from evaluation.
• Figure 5: PGV and $T_{PGV}$ predictions for earthquakes located outside the source distribution used during training. (a) Event at point A (see Figure \ref{['fig:overview']}(b)), located north of the training region along the Hayward Fault. (b) Event at point B (see Figure \ref{['fig:overview']}(b)), located on the San Andreas Fault. For both events: (left) PGV map and (right) $T_{PGV}$ map, with (top) ground truth and (bottom) WaveCastNet predictions.