Hold-out validation results of PRISMA, a single-station model of basin-wide seismic intensities in Mexico City
Year: 2026
Type: article
Abstract: This dataset contains the record-by-record hold-out validation results of PRISMA, a machine-learning model that predicts seismic intensities across the Mexico City basin from the record of a single reference station (CU, Ciudad Universitaria), together with a Python script that recomputes every validation score reported in the associated paper (Cruz-Atienza et al., Earthquake Spectra, submitted). The results cover 2,893 strong-motion records of 85 earthquakes (Mw 5.1–8.2, 1985–2026) at the basin stations. They include all validation tests of the paper: each earthquake held out of training and predicted from its CU record alone; each station held out and compared with the interpolated mean ratios of its neighbors; station blocks and earthquake groups held out together; the assimilation of basin records by kriging and by retraining, also as a function of the number of records assimilated; extrapolation to earthquake families absent from training; the learning curve; and the effect of model complexity. Each file gives, for every record and intensity measure (PSA RotD50 at 25 periods between 0.2 and 5 s, PGA, PGV, CAV, Arias intensity and significant durations), the observed and predicted natural-log ratios between the station and CU. The script compute_scores.py requires only Python 3 with numpy and pandas. It reproduces the scores of Section 4, Table 1, Figures 3, 4, 7 and 9, Text S4 and Tables S5–S6 of the paper. The file README.txt describes each file and column. The archive does not include the strong-motion records, which are available from the operating institutions (RACM-CIRES, Instituto de Ingeniería UNAM and Servicio Sismológico Nacional). Nor does it include the trained models, the site model, the station coordinates, the prediction code or any building-level damage data, because they are part of an operational emergency-response system for Mexico City or are unpublished.
Author: Cruz-Atienza, Víctor M.
Institution: Instituto de Geofísica
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235184162
