Speaker
Description
Seismic forecasting is increasingly being approached as a sequence modeling problem, where modern architectures can directly learn the temporal dynamics of ground motion from raw waveform data. In this work, I introduce a transformer-based framework that treats seismic signals as autoregressive token sequences, enabling short-horizon forecasting. Continuous waveforms are converted into sequences via a simple patch-based tokenization, where fixed-length segments act as tokens, balancing computational efficiency with preservation of local waveform physics.
This approach is motivated by the Einstein Telescope, where low-frequency seismic noise is a dominant limitation. Reliable, low-latency forecasts enable active noise cancellation strategies that anticipate incoming wavefields before they couple into the interferometer.
I will show that transformer models operating on tokenized waveforms can learn meaningful spatio-temporal structure and discuss their extension toward a seismic foundation model.