BLAST-BNN Simulation-to-Field Transfer v1.0.0

Abstract

Reproducibility package for the study “Simulation-to-Field Transfer and Domain-Shift Assessment of Numerical-Data-Trained Models for Blast-Induced Peak Particle Velocity.” This archive contains the complete analysis script, numerical database, external field observations, frozen reference predictions used for exact manuscript reproduction, Python requirements, repository documentation, and software license. The study evaluates zero-shot simulation-to-field transfer of PPV models trained on numerical data only. External field observations are used exclusively for evaluation and are not used for model training, preprocessing, hyperparameter tuning, model selection, domain-threshold definition, or prediction-interval calibration. The reduced QD-BNN uses charge weight (Q) and blast-to-monitoring distance (D) as inputs. The field analysis should therefore be interpreted as a simulation-to-field transfer and domain-shift assessment, not as field validation of the original 12-input BLAST-BNN framework. Reference mode reproduces the locked manuscript results, including the 1080/231/232 numerical split, 46 external field observations, primary Q-D support threshold of 0.146845716514, 8 supported field cases, 38 sparse/OOD field cases, 94.40% numerical-test interval coverage, and 54.35% field interval coverage.