The refinement of acquired well logs has traditionally relied on predefined rock physics models, albeit with their inherent limitations and assumptions. As an alternative, effective yet less explicit machine learning (ML) techniques have emerged. The integration of these two methodologies presents a promising new avenue. In our study, we used four ML algorithms: Random Forests (RF), Gradient Boosting Decision Trees (GBDT), Multilayer Perceptrons (MLP), and Linear Regression (LR), to predict porosity and clay volume fraction from well logs. Throughout the entire workflow, from feature engineering to outcome interpretation, our predictions are guided by rock physics principles, particularly the Gardner relations and the Larionov relations. Remarkably, while the predictions themselves are satisfactory, SHapley Additive exPlanations (SHAP) analysis uncovers consistent patterns across the four algorithms, irrespective of their distinct underlying structures. By juxtaposing the SHAP explanations with rock physics concepts, we discover that all four algorithms align closely with rock physics principles, adhering to its cause–effect relationships. Nonetheless, even after intentionally excluding crucial controlling input features that would inherently compromise prediction accuracy, all four ML algorithms and the SHAP analysis continue to operate, albeit in a manner that seems irrational and starkly contradicts the fundamental principles of rock physics. This integration strategy facilitates a transition from solely mathematical explanations to a more philosophical interpretation of ML-based predictions, effectively dismantling the traditional black box nature of these ML models.

In petroleum exploration and production, extracting petrophysical parameters of oil reservoirs from 3D seismic elastic data poses significant challenges. Applying machine-learning (ML) algorithms to this issue promises much, but rigorous feasibility evaluation in advance is crucial. More significantly, a comprehensive mathematical and logical justification is imperative to illuminate the “black box” nature of ML, thereby bridging existing gaps in explanation for enhanced trust and insights. In this study, the input elastic data for random forests (including P- and S-wave velocities as well as density) is derived from the Raymer-Dvorkin-Voigt rock-physics model with known petrophysical parameters. Subsequently, noise or filtering is applied to the input data to emulate field application conditions. The modeled data, devoid of confounders, facilitate unambiguous causal interpretation. The algorithm demonstrates efficient and stable predictions for porosity, followed by clay volume, whereas water saturation predictions are less accurate. Subsequently, Shapley additive explanations are used to mathematically assess the relative importance of each input feature in determining the outcomes, thereby elucidating the algorithm’s functionality. Furthermore, the causality is understood by referencing rock physics as the foundational assumption of truth. The results indicate illogical behavior in association-oriented ML, contradicting rock physics and logical principles. Overall, ML for petrophysical prediction is effective but constrained by several factors, such as the problematic assumption of independent and identical distributed data, the unsatisfactory quality of data, and the complicated nature of the underlying physics. Our framework establishes a solid foundation for subsequent inquiries aimed at elucidating the explanation and causality behind ML-based petrophysical predictions.