SCEC Award Number 24105 View PDF
Proposal Category Individual Research Project (Single Investigator / Institution)
Proposal Title An AI-based Approach for Seismic Wavefield Reconstruction from Sparse Field Observations in San Francisco Bay Area
Investigator(s)
Name Organization
Nils Benjamin Erichson International Computer Science Institute Rie Nakata International Computer Science Institute
SCEC Milestones C2-2, A1-1, A1-2, C1,2,3-1, C1-2 SCEC Groups RC, GM, Seismology
Report Due Date 03/15/2025 Date Report Submitted 05/19/2025
Project Abstract
 We develop and evaluate two waveform interpolation methods to generate gridded training datasets for WaveCastNet, a machine learning framework designed for earthquake early warning (EEW). WaveCastNet predicts future ground motions from early waveform segments recorded at sparse stations, eliminating the need for phase picks or magnitude estimates. To support training, we implement two strategies: (1) a physics-based method using phase-guided interpolation and dynamic time warping (DTW), and (2) a machine learning approach using the Swin Transformer, a hierarchical vision Transformer adapted for seismic data. The physics-based method estimates P-wave arrivals using a hyperbolic moveout model and applies DTW to align waveforms across stations, enabling interpolation of coherent wavefields. This approach is tested on synthetic data for the 2018 M4.4 Berkeley earthquake, showing good agreement with ground truth waveforms but limitations in areas with sparse station coverage. The ML-based method uses observed data, station locations, and initial interpolations to reconstruct wavefields via deep feature extraction with Swin Transformers. Trained on 35 earthquakes recorded at a dense array in Milford, Utah, the model successfully recovers both P- and S-wave arrivals at frequencies up to 20 Hz, outperforming the initial interpolation inputs. Both methods produce realistic wavefields and demonstrate strong potential for training data augmentation. However, their accuracy decreases with sparse station spacing. Future work will focus on improving these methods and evaluating their applicability to real seismic data for robust EEW deployment.
Intellectual Merit This project developed both physics-based and machine learning (ML) approaches to interpolate sparsely sampled data and reconstruct densely sampled seismic wavefields from field observations. These interpolated datasets can serve as input for source-to-site (S2S) ground motion modeling and tomographic inversions for subsurface velocity model development. Our focus area is the San Francisco Bay Area (SFBA), one of the most densely populated metropolitan regions in the U.S. Its high seismic hazard, primarily due to the Hayward Fault and complex geological structures, necessitates an advanced early warning system that accounts for highly heterogeneous wave propagation.
Broader Impacts This project contributes to advancing earthquake early warning capabilities, ground motion model building, and seismic velocity model estimations by reconstructing dense seismic wavefields from sparse observations, supporting faster and more accurate alerts in high-risk areas like the SFBA. These improvements contribute to public safety, disaster resilience, and equitable access to hazard information. The project also fostered interdisciplinary collaboration between seismology and machine learning, promoting innovation across fields. Additionally, it supported workforce development by training one postdoctoral researcher and one graduate student in seismological and ML techniques, equipping the next generation of scientists to address critical societal challenges through cross-disciplinary approaches.
Project Participants Rie Nakata (ICSI) Benjamin Erichson (ICSI) Nori Nakata (ICSI) Zhengfa Bi (LBL) Garry Gao (ICSI)
Exemplary Figure Figure 2: Data augmentation results. (a) Map view showing station locations (blue triangles), the earthquake
epicenter (yellow star), active fault traces (black lines), and the interpolation domain (black rectangles).
(b) Snapshots at 2.1 seconds for the (left) simulated ground truth and (right) interpolated result,
with stations (white triangles) and the earthquake epicenter (yellow star) indicated. The cyan and yellow
lines mark the locations of the arrays used in panels (c) and (d). (c–d) Waveforms along the arrays in the
(c) east–west (EW) and (d) north–south (NS) directions for the (left) ground truth and (right) interpolated
results.
Linked Publications

Add missing publication or edit citation shown. Enter the SCEC project ID to link publication.