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.

Traditional geological mapping is often time-consuming, labor-intensive, and restricted by rugged terrain. This study addresses these challenges by proposing a novel methodology for automated lithological identification in the Ququleke area of the eastern Kunlun Mountains, which pioneers the integration of portable UAV oblique photogrammetry with a Coordinate Attention-enhanced DeepLabV3+ (CA-DeepLabV3+) semantic segmentation framework for geological mapping. Using a DJI Mavic 3M quadcopter, high-resolution oblique photogrammetric orthophotos were captured to build a pixel-level lithology dataset containing four classes: sandstone, diorite, marble, and Quaternary sediments. The CA-DeepLabV3+ model, adapted from the DeepLabV3+ encoder–decoder framework, integrates a lightweight MobileNetV2 backbone and a Coordinate Attention mechanism to strengthen spatial position encoding and fine-scale feature extraction, crucial for detailed lithological discrimination. Experimental evaluation demonstrates that the proposed model achieves an overall accuracy of 97.95%, mean accuracy of 97.80%, and mean intersection over union of 95.71%, representing a 5.48% improvement in mean intersection over union (mIoU) over the standard DeepLabV3+. These results indicate that combining UAV oblique photogrammetry with the CA-DeepLabV3+ network enables accurate lithological mapping in complex terrains. The proposed method provides an efficient and scalable solution for geological mapping and mineral resource exploration, highlighting the potential of low-altitude UAV remote sensing for field-based geological investigations.

The elastic properties of argillaceous sandstones are significantly controlled, however, by the perplexing distribution of dispersed, cemented and matrix (including structural and layered) clay. The corresponding rock physics models are established to investigate the influence of different clay distribution on the elastic properties of sandstone. The rock physics modelling and laboratory experimental results exhibit that the higher the content of the matrix clay, the lower the elastic wave velocity. The dispersed and cemented clay increases the sandstone's velocity by reducing the porosity; the increase of dispersed clay only causes a slight increase in the elastic wave velocity in sandstone. In contrast, a small amount of cemented clay can significantly increase the velocity in sandstone. Based on these understandings, we construct a cross-plot of P-wave velocity and clay content as a template to diagnose clay distribution. Based on the rock physics model and empirical knowledge, we divide the template into four zones, namely, matrix, dispersed, cemented and mixed clay distribution zones. Then, we use the published experimental data and numerical modelling data to validate the template. Based on diagnosing the clay distribution, we introduce how to use the diagnostic results of the template to select a suitable rock physics model for argillaceous sandstone in Shengli Oil Field, East China. The selected rock physics model, guided by diagnosing the clay distribution, predicts P- and S-wave velocity very well. The proposed rock physics template for diagnosing clay distribution also has potential application in well-logging data interpretation, rock physics modelling and reservoir characterization.

In petroleum exploration and production, accurate reservoir characterization and seismic modeling depend on linking macroscale seismic data with microscale reservoir properties. Prior research has predominantly concentrated on forward modeling or inversion processes in isolation. As a result, a comprehensive multiscale framework that seamlessly integrates both approaches remains absent. In this study, to address this gap, a machine learning (ML)-based petrophysical inversion method was developed. This method was integrated into a unified multiscale workflow by combining rock physics modeling, seismic modeling, and seismic inversion techniques. This study was based on synthetic data. Initially, predefined reservoir petrophysical parameters were used as inputs to rock physics equations to forward model elastic parameters. Following this, seismic forward modeling was performed to generate seismic amplitude-versus-offset (AVO) data. Subsequently, seismic AVO inversion was carried out to recover elastic parameters from the AVO data, which were then converted into reservoir petrophysical parameters using ML techniques. As a result, forward modeling revealed that porosity (𝜙) significantly affects seismic AVO responses, whereas clay volume (C) and water saturation (S_w) had minimal impact. Conversely, seismic AVO inversion, constrained by wavelet effects and input uncertainties, produced a filtered and biased representation of elastic parameters. This compromised the accuracy of subsequent ML-based petrophysical inversion, particularly for S_w in oil reservoirs and C in gas reservoirs. Consequently, 𝜙 inversions demonstrated high reliability, whereas S_w and C predictions showed greater uncertainty. Therefore, by integrating petrophysics with ML, the proposed methodology effectively bridges micro-properties with macro-seismic signals, offering a precise and unified multiscale approach for quantitative seismic modeling and reservoir characterization. Furthermore, the present study capitalizes on rock-physics-generated data, providing ground truth and enabling rigorous, multiscale uncertainty analysis beyond the capabilities of isolated methods. Additionally, it can be seamlessly applied to real-field data. Finally, a comparative discussion between ML-based and physics-based methodologies was performed, leading to the recommendation of implementing explainable artificial intelligence (XAI) for improved interpretability and prediction performance.