Fig. 9: Performance comparison between sequence-to-seqence frameworks with different recurrent cells and state-of-the-art transformers in both in-domain and out-of-domain settings. | Nature Communications

Fig. 9: Performance comparison between sequence-to-seqence frameworks with different recurrent cells and state-of-the-art transformers in both in-domain and out-of-domain settings.

From: Rapid wavefield forecasting for earthquake early warning via deep sequence to sequence learning

Fig. 9: Performance comparison between sequence-to-seqence frameworks with different recurrent cells and state-of-the-art transformers in both in-domain and out-of-domain settings.

The evaluation metric used is RFNE (Relative Frobenius Norm Error). a In-domain forecasting performance on point-source earthquakes as a function of model parameter size. b Generalization performance as a function of earthquake magnitude. All models are trained only on point-source earthquakes. Colors indicate model type and parameter count (in millions, consistent with (a)): yellow for Swin Transformer53,54 (13.72 million parameters), orange for Swin Transformer* (24.27 million parameters), light blue for Time-S-Former55(10.21 million parameters), and dark blue for Time-S-Former* (33.82 million parameters). *indicates larger parameter volume. While larger vision transformers may achieve higher accuracy on in-domain tasks, WaveCastNet generalizes best to domain-shifted settings. Detailed configurations and evaluation metrics for all baselines are provided in Supplementary Table D.1.

Back to article page