A machine learning diagnostic test for nonstationary ETAS aftershock models

Nonstationary extensions of the Epidemic Type Aftershock Sequence (ETAS) model, in which the background rate and triggering productivity vary in time, are increasingly used to infer transient post-seismic processes from earthquake catalogs. Their flexible parameterization, however, raises an identifiability problem: an improved statistical fit may stem from parameter trade-offs rather than from a genuine change in the seismic regime. We present a diagnostic that pairs penalized nonstationary ETAS inference with an independent, model-agnostic anomaly-detection analysis, and ask whether inferred parameter transients are accompanied by observable changes in catalog properties. We apply it to the aftershock sequence of the 2025 M w 7.1 Dingri earthquake in southern Tibet, for which independent geodetic and mechanical constraints are available. An empirical-Bayes nonstationary ETAS inversion, with smoothing weights selected by ABIC, yields time-varying background rate and triggering productivity, with the strongest variations about 8–10 days after the mainshock. Using an Isolation Forest applied to temporal, magnitude, depth, and spatial features, we find no coherent anomaly during this interval. Semi-synthetic experiments that perturb only the temporal occurrence rate, consistent with the temporal ETAS formulation, together with feature-ablation tests, show that the detector responds primarily to temporal-rate changes and recovers the injected transient when present. The absence of a corresponding anomaly in the real catalog indicates that the inferred parameter variations should be interpreted with caution. More broadly, the results show the value of validating nonstationary ETAS inferences against independent, model-agnostic diagnostics.