Investigation of rock-mass fragmentation by explosive charges using interpretable machine learning and literature-derived air-deck and related decked-charge blasting data

Abstract Decked-charge blasting includes axially discontinuous or decoupled charge configurations, with air-deck blasting as one subtype. Here, air-deck denotes an intentional air or gaseous void, whereas decked-charge includes air, water, expanded-polystyrene (EPS), and polyethylene (PE) decks. Published fragmentation results remain inconsistent because rock conditions, blast geometry, deck design, response definition, and reporting differ across studies. This study develops a literature-derived, source-aware evidence-synthesis and exploratory interpretable-modelling framework. A dataset of 42 observations and 114 fields from six sources was reconstructed; 30 observations from four sources were regression-eligible, comprising 20 air-deck, two water-deck, one EPS-deck, one PE-deck, and six conventional/no-deck cases. Random Forest achieved leave-one-out MAE = 82.29 mm, RMSE = 126.50 mm, and R 2 = 0.391. Bayesian optimization improved tuning-stage performance, but nested leave-one-source-out validation remained weak. Independent external validation on 31 field cases from three unseen publications yielded a best pooled result of MAE = 134.64 mm, RMSE = 158.27 mm, and R 2 = − 0.194; all evaluated models had negative external R 2 . Explainability identified response family, rock strength, burden-to-diameter ratio, hole depth, observation scale, hole diameter, specific charge, stemming, bench-height-to-burden ratio, and deck position as influential. Sensitivity analysis showed that within-dataset associations were not driven solely by response family, while cross-source transfer remained limited by source and response-definition heterogeneity. Source-level comparison found smaller fragmentation in 11 of 18 paired cases (mean − 21.91%; median − 31.48%). Air-decking should therefore be treated as a site- and configuration-dependent option requiring local verification, and the framework as an evidence-synthesis tool rather than a direct blast-design predictor.