Extended Data Fig. 8: Synthetic test with Tohoku STF.
From: Instantaneous tracking of earthquake growth with elastogravity signals
Density plot of the predictions on 1,000 examples obtained by combining different noise recordings (randomly extracted from the test set) with PEGS waveforms, from a synthetic source that mimics that of the Tohoku-Oki earthquake. The median, Q1–Q3 interquartile range and the 5th–95th percentiles of the distribution are reported with solid, dotted and dashed red lines, respectively. The synthetic Tohoku-like source is obtained from the ‘true’ STF (orange line, same as in Fig. 3 with ±0.3 Mw indicated by dashed lines), the hypocentre location (Extended Data Fig. 10a), and the following values for the strike, dip and rake: 193.0°, 8.9° and 78.4° (USGS catalogue). The workflow described in Methods is used to combine synthetics and noise. PEGSNet prediction for real Tohoku data is reported for comparison (blue line, same as in Fig. 3). Even if it has been trained on synthetic data plus noise, PEGSNet is able to generalize well to real data (the blue line is within the expected variability of the predictions obtained on synthetics). In the first 50 s after origin, the variability of the predictions is high and strongly affected by noise (similar to Fig. 2c), whereas after 50 s, predictions have similar Mw(t) values and therefore are more robust. Note that for the lowest noise conditions, faster PEGSNet response is virtually possible, as early as about 30 s after origin (as indicated by the 95th percentile of the distribution).