Title: Transformer-Based Architecture for Data-Driven Forecasting of Three-Component Seismograms

Abstract:

Predicting seismic waveforms in regions where data has not yet been recorded is a significant challenge, largely because seismic wave propagation exhibits complex, nonlinear, and dispersive characteristics across multiple scales. To address this, we present \textsc{SeismoGPT}, an autoregressive model built on transformer architecture that forecasts three-component seismic waveforms directly within the time domain. We frame the forecasting task as a physically constrained continuation problem: the model ingests waveform context starting from the P-wave arrival and extending past the S-wave arrival, then recursively generates future motion without relying on ground-truth samples.

We evaluated the model using synthetic seismograms characterized by source depths ranging from 5 to 100 km, epicentral distances between 10 and 90 degrees, and moment magnitudes ($M_w$) from 3 to 7. To isolate the impact of context length from prediction horizon, we established three distinct evaluation configurations. These utilized a distance-normalized context ratio alongside fixed prediction horizons of 120 and 240 seconds. The model consistently achieved median normalized cross-correlation scores exceeding 0.93 across all scenarios.

Examination of representative forecasts indicates that successful predictions maintain both phase coherence and the distribution of spectral energy. In instances where the model failed, the primary cause was gradual phase drift during the autoregressive rollout process, rather than the generation of unphysical signals. These findings suggest that transformer-based sequence models are capable of learning stable dynamical continuations of seismic wavefields. This underscores the promise of foundation-model methodologies for physics-driven time-series forecasting. Potential applications for this approach include seismic warning systems and hazard mitigation strategies, with specific relevance to next-generation gravitational-wave observatories like the Einstein Telescope.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC