An exploratory machine learning framework developed to evaluate rock fragmentation across published air-deck and decked-charge blasting studies demonstrated that predictive models trained on pooled literature data fail to generalize reliably to new mine sites and independent publications. Investigating rock breakage produced by explosive charges, a research team reconstructed forty-two case-level observations with one hundred fourteen design, explosive-use, rock-property, and performance fields extracted from six published sources. Among these extracted records, thirty observations from four publications met strict eligibility criteria for regression modelling of fragmentation particle size. In benchmark evaluations, a supervised Random Forest algorithm achieved a leave-one-out cross-validation mean absolute error of 82.29 millimetres, a root mean squared error of 126.50 millimetres, and a coefficient of determination of 0.391. However, when evaluated under leave-one-source-out cross-validation where an entire literature source is held out from training, the Random Forest model performance deteriorated markedly to a mean absolute error of 175.41 millimetres, a root mean squared error of 208.08 millimetres, and a negative coefficient of determination of minus 0.648. This marked drop in predictive capability demonstrates that statistical associations learned from compiled literature reflect dataset-specific conditions rather than transferable physical rules for engineering blast design.
Drilling and blasting operations represent critical initial stages across surface mining, quarrying, and civil tunneling projects, directly dictating downstream energy consumption and operational costs during subsequent mucking, hauling, crushing, and mineral processing. In conventional bench blasting, explosive energy distribution is frequently non-uniform throughout the blasthole column, generating excessive boulder-sized oversize blocks near the unconfined upper bench crest alongside unnecessary ultrafine pulverized rock adjacent to the concentrated explosive charge. Decked-charge blasting encompasses axially discontinuous or decoupled charge configurations designed to distribute explosive shock energy and high-pressure expansion gases more evenly along the borehole length. Air-deck blasting represents a specific variant where the continuous explosive column is deliberately interrupted by an intentional air or gaseous void, whereas decked charges more broadly encompass inert intervening media such as water, expanded polystyrene, or solid polyethylene plugs. Despite substantial industrial interest and documented physics suggesting air gaps can suppress local over-crushing and improve top-bench breakage, published fragmentation results across international literature remain contradictory due to unstandardized blasthole geometries, differing rock mechanics, varied deck placements, and conflicting measurement methods.
To rigorously examine these conflicting empirical claims, the investigators established a traceable, source-aware evidence-synthesis architecture rather than presuming published studies describe identical physical interventions. The development evidence base gathered forty-two case-level blasting records from six peer-reviewed publications spanning 1999 to 2025 across major citation platforms. From this curated repository, thirty observations derived from four primary publications provided direct row-level numeric fragmentation sizes alongside corresponding engineering variables suitable for supervised regression. The resulting regression-eligible dataset comprised twenty air-deck configurations, two water-deck cases, one expanded polystyrene deck, one polyethylene deck, and six conventional fully coupled control blasts lacking axial decks. The reported target fragmentation sizes spanned from 20.60 millimetres to 641.00 millimetres with an overall median of 217.90 millimetres. The predictive modelling pipeline incorporated twenty-three input features, encompassing eighteen continuous numerical variables and five categorical descriptors capturing hole diameter, burden, spacing, stemming length, bench height, hole depth, rock compressive strength, specific charge, explosive density, deck count, deck position, deck medium, and dataset reporting structure.
Supervised regression benchmarking examined five alternative algorithmic configurations: a baseline median predictor, Ridge Regression, Random Forest, Extra Trees, and Histogram Gradient Boosting. Within-fold median imputation for missing continuous variables and mode imputation followed by one-hot encoding for categorical factors were executed strictly inside cross-validation splits to prevent information leakage from test records. While Ridge Regression and Extra Trees yielded leave-one-out mean absolute errors of 84.59 millimetres and 88.35 millimetres respectively, Random Forest attained the lowest error among predefined models and was selected as the focal framework. To explore whether predefined hyperparameter constraints restricted predictive accuracy, the researchers implemented secondary Bayesian hyperparameter optimization using the tree-structured Parzen estimator in Optuna across thirty trials per architecture. Hyperparameter tuning enhanced within-dataset fit across all candidate families, with optimized Histogram Gradient Boosting lowering tuning-stage leave-one-out mean absolute error to 74.84 millimetres with a root mean squared error of 115.49 millimetres and a coefficient of determination of 0.492. Nevertheless, when subjected to rigorous nested leave-one-source-out cross-validation where hyperparameters were selected exclusively on inner folds, optimized models continued to yield negative coefficients of determination, confirming that algorithmic tuning cannot bridge profound structural disparities between divergent literature sources.
Model explainability diagnostics illuminated how algorithmic predictions relied on both physical blasting variables and reporting-level artifacts embedded within published studies. Permutation feature importance analysis revealed that the categorical response family had the largest single impact on model accuracy, increasing mean absolute error by 27.11 millimetres when permuted, followed by rock-strength classification with 16.43 millimetres, and burden-to-diameter ratio with 14.68 millimetres. Aggregated impurity-based tree importance exhibited an identical structural prioritization, identifying response family, rock strength, burden-to-diameter ratio, hole depth, observation scale, hole diameter, specific charge, stemming length, bench-height-to-burden ratio, and deck position as influential inputs. The elevated importance assigned to response family stems from the fact that sixteen regression records documented mean fragment size while fourteen documented median fragment size x50, with response definitions perfectly nested within literature sources. Bootstrap resampling across five hundred source-stratified replicates yielded a median Spearman rank correlation of 0.775, confirming that while an upper tier of influential descriptors consistently governed predictions, the precise ordinal hierarchy of individual engineering parameters fluctuated significantly depending on case composition.
To inspect model behavior without extrapolating into physically implausible configurations, the authors evaluated constraint-aware one-dimensional response profiles and two-dimensional response surfaces restricted strictly to empirically supported input domains. During parameter variation, algebraically linked geometric descriptors were dynamically updated to preserve physical definitions, including burden recalculations when shifting burden-to-diameter ratios and stemming adjustments linked to blasthole dimensions. Numerical profiles illustrated that fitted fragment sizes generally decreased across supported burden-to-diameter ranges while expanding sharply at deeper hole depths. For bivariate response surfaces, predictions outside the convex hull of actually observed pairwise data points were masked, revealing severe data sparsity; joint observations of burden-to-diameter ratio with hole diameter covered only 57.5 percent of their bounding rectangular space, burden-to-diameter with specific charge covered 50.4 percent across eleven cases, and hole depth with specific charge occupied only 60.1 percent across merely six observations. Categorical profiles further verified that records classified under the unknown rock-strength category produced larger predicted fragment dimensions, reflecting incomplete field reporting rather than rock mechanics.
Separately from machine learning predictions, the authors conducted paired source-level control comparisons to assess how decked charges altered rock breakage relative to specific baseline blasts executed within identical publications. Across eighteen primary treatment-control pairings containing row-level fragmentation sizes, eleven decked cases achieved finer fragmentation than their source-defined control blasts, whereas seven produced coarser fragment dimensions. The pooled mean fragmentation change was minus 21.91 percent with a wide ninety-five percent hierarchical bootstrap uncertainty interval spanning from minus 56.42 percent to positive 22.94 percent, while the pooled median change was minus 31.48 percent with an interval from minus 61.72 percent to positive 31.68 percent. Because both bootstrap intervals encompass zero, the literature does not establish a statistically consistent global fragmentation reduction. Furthermore, source-stratified contrasts showed pronounced divergence: Hong and colleagues exhibited a median fragmentation reduction of minus 53.77 percent across ten comparisons, Monjezi and co-authors recorded a median increase of positive 31.20 percent across seven comparisons, and Kabwe and Banda documented a single contrast showing a minus 31.17 percent reduction. Sensitivity audits indicated that excluding Hong and colleagues shifted the pooled mean to positive 9.34 percent, demonstrating that directional conclusions remain highly dependent on source selection.
To provide a definitive test of operational transferability, the researchers evaluated the frozen machine learning models against an entirely independent external validation dataset comprising thirty-one field blasting cases compiled from three previously unexamined publications by Shehu, Saqib, and Balamadeswaran. Across these thirty-one independent field blasts, which encompassed twenty-one mean fragment size observations and ten x50 size records, all evaluated models failed to attain positive predictive skill. The tuned Histogram Gradient Boosting model achieved the lowest pooled error with a mean absolute error of 134.64 millimetres, a root mean squared error of 158.27 millimetres, and a negative coefficient of determination of minus 0.194. The baseline Random Forest yielded a mean absolute error of 146.39 millimetres and a coefficient of determination of minus 0.459, whereas the optimized Extra Trees model produced a mean absolute error of 165.25 millimetres and a coefficient of determination of minus 0.997.
A critical technical barrier preventing model transferability is severe covariate shift and unharmonized measurement protocols between distinct operations. In the external dataset, multiple blast parameters extracted from Shehu and colleagues extended far beyond the domain ranges encountered during model development, including extreme specific charge values, atypical burden-to-diameter ratios, and deeper hole dimensions. Beyond geometric discrepancies, measurement techniques diverged fundamentally across studies: Hong and co-authors utilized two-dimensional digital photography analyzed through ImageJ fitted to generalized extreme value distributions; Monjezi and colleagues implemented Split-Desktop software coupled with Kuz-Ram calculations; Jhanwar and co-authors relied on physical manual measurement of approximately two hundred rock fragments combined with visual muckpile inspections; and Kabwe and Banda applied Split-Desktop digital image processing.
These comprehensive findings demonstrate that machine learning models fitted to heterogeneous published blasting literature cannot serve as off-the-shelf predictive tools for operational blast engineering or automated parameter selection. While algorithmic synthesis effectively identifies recurring physical associations involving blast geometry, explosive intensity, and rock mass competence, it also exposes how easily reporting deficits and source-level clustering can be mistaken for genuine physical rules. Mining engineers and blasting practitioners should treat air-decking and alternative decked-charge designs as highly site-specific techniques that demand rigorous local calibration against dedicated baseline blasts. Future advances in data-driven rock fragmentation modelling require standardized multi-site prospective field monitoring, uniform muckpile granulometry protocols, fully disclosed borehole and deck geometries, and deliberately overlapping operational domains to separate genuine rock breakage mechanics from the idiosyncrasies of individual mining studies.
Subject of Research: Interpretable machine learning synthesis and cross-source validation of rock-mass fragmentation in decked-charge and air-deck blasting
Article Title: Investigation of rock-mass fragmentation by explosive charges using interpretable machine learning and literature-derived air-deck and related decked-charge blasting data
Article References: Ochilov, S., Urunov, B., Azamat, U., Yaxshiyev, B., Karamanov, A., Kushnazorov, I., Raimkulova, S., Dexkonova, N., Orzugulkhon, A., & Javohir, R. (2026). Investigation of rock-mass fragmentation by explosive charges using interpretable machine learning and literature-derived air-deck and related decked-charge blasting data. Journal of Engineering and Applied Science, 73(1), Article 438. https://doi.org/10.1186/s44147-026-01305-y
Image Credits: AI Generated
DOI: 10.1186/s44147-026-01305-y
Keywords: rock fragmentation, air-deck blasting, decked-charge blasting, machine learning, Random Forest, blast design, mining engineering, leave-one-source-out validation, external validation, granulometry, mean fragment size, x50 median size, specific charge, burden-to-diameter ratio, Bayesian optimization, explainable AI, rock mechanics
News Source: Teresa Odom. (November 20, 2026). Machine Learning Synthesizes Blasting Data but Fails Cross-Source Fragmentation Prediction. Scienmag.



