Answers are grounded in this paper and cite their sources.
Onsite earthquake early warning (EEW) must predict shaking from the first seconds at a single station. We present a sourceagnostic framework for Indonesia’s Java–Sumatra region using only waveforms and site V S30 : a notice of damaging shaking from the first 0.5 s of the P wave (~1.9 s available), a confirmation and routing at 3 s, and the 5%-damped horizontal geometric-mean response spectrum S a (T) at 102 periods at a window chosen by risk-gated routing (Intensity-Driven Adaptive P-wave Time Windowing, IDA-PTW). On 39,434 records (1,135 earthquakes; 6,038 prospective holdout), the 3 s classifier reaches ROC-AUC 0.9718 for damaging shaking (III+, PGA > 16 gal) at a false-alarm rate of 0.63%, and 0.9777 when folds are grouped by station. The 0.5 s notice gives 9 false notices per station-day on noise and suits only reversible actions. For damaging records, Adaptive-PTW raises the spectral R 2 log of the fixed 3 s window from 0.23 to 0.36 and cuts the amplitude bias from −0.241 to −0.193 dex; fixed 8 s and 10 s windows remain more accurate (0.63, 0.68) because 10% of damaging records are not escalated. On the holdout, skill is stable for all records (0.785) but not for damaging ones (0.002).
No takes yet. Share an insight, caveat, or question.
Nugraha et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: