A Brazilian seismic acceleration database and multicriteria GMPE evaluation for national PSHA

Ground-Motion Prediction Equations (GMPE) are a fundamental component of Probabilistic Seismic Hazard Analysis (PSHA). In low-seismicity regions such as Brazil, the lack of locally calibrated GMPE has led to the adoption of models developed for tectonically analogous regions, particularly Central and Eastern North America (CENA). Previous Brazilian PSHA studies based on foreign GMPE produced high peak ground acceleration (PGA) estimates, raising concerns regarding their engineering applicability. In this study, a database comprising approximately 150 PGA records from 27 regional earthquakes (2.8 ≤ ≤ Mw ≤ ≤ 5.6), combined with site characterization surveys measuring V S30 S 30 , is used to evaluate a pool of 48 candidate GMPE spanning five families: the NGA-East final and seed models, and models developed for the United Kingdom, Australia, and Canada. Events below the Mw 4.0 validity floor of most GMPE are evaluated at their true recorded magnitude, with each candidate independently extrapolated downward using the magnitude-scaling shape of Yenier & Atkinson (2015). Candidates are screened in two stages and assessed at two distance ranges (R ᵣᵤₚ rup ≤ ≤ 100 and ≤ ≤ 400 km) through a set of complementary metrics, including normalized residuals (Z-scores), root-mean-square error (RMS), magnitude- and distance-binned RMS, log-likelihood (LLH), the stochastic-area metric (SA), and a random-effects decomposition of residual variance into inter-event, site-to-site, and remaining components. These metrics are reported side by side and discussed qualitatively, rather than combined into a composite score. Most candidates overpredict Brazilian PGA on average at both ranges, plausibly related to differences in crustal attenuation, stress-drop, and focal-depth distributions between CENA and the Brazilian stable platform, although these hypotheses are not directly tested by the present dataset. Three models stand out on complementary grounds and are suggested as a starting point for the epistemic-uncertainty branches of future Brazilian PSHA logic trees: nga_east_usgs_13, with the lowest magnitude-binned RMS and site-to-site variability at 100 km; nga_east_Graizer, the only shortlisted model with a bias not statistically distinguishable from zero at 400 km; and Allen2012_SS14, with the lowest raw RMS at both ranges and attenuation assumptions developed independently of the NGA-East framework. Their residual trends with distance run in opposite directions, so that together they bracket the observed attenuation rather than sharing a common bias. The compiled dataset is made publicly available, and the proposed evaluation framework offers a reproducible methodology transferable to other low-seismicity intraplate settings.