Abstract
Traditional rock typing in reservoir engineering relies on destructive, costly, and slow macroscopic core measurements. We propose an explainable machine-learning framework, of which the central contribution is a triple-anchor cross-paradigm audit that confronts the supervised model’s SHAP attribution with three closed-form permeability laws of increasing feature breadth: classical Kozeny–Carman kKC (2 features), pore-network Hagen–Poiseuille kHP (4 features), and Koponen-style modified Kozeny–Carman kK (6 features). Using the eleven-sandstone μ-CT collection of Neumann et al., 24,167 sub-volumes were extracted and 12 pore-network descriptors were computed per sub-volume via pnextract. An XGBoost classifier trained under spatial-block 5-fold cross-validation that respects the tiled-window structure of the μ-CT data achieved an honest accuracy of 91.20%±1.08% (89.38% on a fully external held-out spatial block; 93.98% as a random-split upper bound), exposing a ∼2.8 percentage-point leakage bias of conventional random splits. Per-class recall is high for most formations but drops to 0.51 for one member of the genetically related Berea family (Berea Upper Gray), which accounts for the bulk of the residual error. All three permeability laws agree on the hydraulic end-points, with Parker the lowest-permeability member (k̄KC≈1.6 mD) and the high-permeability lobe (Bentheimer, k̄KC≈64.6 mD; Castle Gate; Kirby) at the top of the ordering. The audit returns Spearman ρ=+0.32 against kKC and ρ=−0.08 against kHP, both formally non-significant; with the six-feature Koponen-style modified Kozeny–Carman law, which couples porosity to both count statistics, both mean radii, and the throat length through an explicit surface-area term, the convergence rises to ρ=+0.84 (p=0.0007). The classifier is physics-anchored only when the closed-form law spans the same pore-network statistics it uses, making anchor choice an explicit and testable criterion rather than an unstated assumption. The same six-feature Koponen anchor that aligns with the model also reproduces the independently measured laboratory permeabilities of the eleven sandstones most closely (Spearman ρ=+0.96 at the formation level), and the convergence pattern persists when the classifier is replaced by a random forest or a neural network, indicating that the audit is a model-agnostic protocol rather than an artifact of XGBoost. This auditable, non-destructive workflow is intended as a fast triage step for CCS site screening and unconventional sandstone characterization; its extension to actual drill cuttings, whose damaged and multi-mineral pore structure differs from intact plugs, is identified as a priority for future validation.