Abstract
Rock mass compactness is critical for geotechnical characterization, but core logging is slow. This study develops an auditable framework to classify rock mass compactness (Compact, Semi-Compact and Friable) from the granulometry of reverse-circulation drill chips. Tyler particle-size distributions from N=60 intervals of an itabirite iron-ore deposit were projected by a supervised autoencoder onto a two-dimensional phase space (Z1,Z2) in which multinomial logistic regression defines closed-form decision boundaries. Over 20 repetitions of leave-one-out cross-validation, the framework achieved 65.4±4.3% accuracy, exceeding linear classifiers trained on the sieve fractions (50.0–56.7%); a Random Forest reached 73.3% but provides no planar representation. Withholding entire drill holes reduced accuracy to 35.0% (baselines 18–43%); so, performance applies to the sampled holes. Y-randomization gave p=0.001, the minimum attainable. In the archived reference model, encoder Jacobians showed that Z1 reflects the depletion of intermediate chips (#28–#14) and Z2 coarse-chip retention (#14 and #6), a reading specific to that model because the axes are not fully reproducible between training runs. The encoder reduces to spreadsheet-implementable matrix operations, and the site-calibrated framework requires validation before transfer to other deposits.
Funding Information
  • Empresa Brasileira de Pesquisa e Inovação Industrial (23109.006004/2025-55)
  • Conselho Nacional de Desenvolvimento Científico e Tecnológico