The central challenge is what geologists call a dual-medium reservoir. In these formations, gas lives in two very different storage systems at once: the tiny pores between sand grains, and the sparse network of natural fractures that cuts across them. Fractures act like high-speed highways for gas flow, while the matrix pores act like vast parking lots. Knowing which system dominates in any given interval of rock is essential for deciding where to drill and how to stimulate a well. But conventional classification methods rely on discrete analytical samples, mostly from core pieces pulled from a handful of wells. Most wells are never cored at all, so point-based analysis simply cannot capture how fractured and matrix-dominated zones alternate continuously down the borehole.

The research team’s first contribution is a revised classification framework, which they describe as a three-major and six-minor scheme, built on pore structure. It separates the Xujiahe sandstones into nine reservoir types spanning matrix-controlled, fractured, and hybrid matrix-fractured varieties. A key analytical device is the fracture identification factor, a quantity derived from well logs that distinguishes true fractured reservoirs from matrix-fractured ones. Because well logging curves are recorded continuously as instruments travel up the borehole, using them as classifier inputs means every centimeter of rock in every well can be typed, not just the rare intervals where core was recovered.

But continuous logging data brought a familiar artificial intelligence problem to the surface: class imbalance. Some reservoir types, particularly the highest-quality fractured varieties, are genuinely rare in the subsurface, so their examples are heavily outnumbered in any training dataset. Machine learning classifiers trained on such lopsided data tend to simply ignore the minority classes, which is precisely the wrong behavior when those rare classes are the exploration targets. To fix this, the team applied ADASYN, the adaptive synthetic sampling technique. ADASYN works by generating synthetic minority-class examples, concentrating its effort near the decision boundaries where classification is hardest. Rather than flooding the dataset with generic duplicates, it teaches the model to pay attention to the ambiguous frontier zones between reservoir types, which the authors found markedly strengthened recognition of the scarce but valuable classes.

With rebalanced data in hand, the researchers then assembled what they call a multi-ensemble learning architecture. Instead of trusting a single algorithm, they trained four different base classifiers: random forest, support vector machine, multilayer perceptron, and XGBoost. Each brings a distinct inductive bias. Random forests average many decision trees trained on random subsets of the data, support vector machines carve maximal-margin boundaries through feature space, neural networks learn flexible nonlinear mappings, and XGBoost builds boosted trees that iteratively correct earlier errors. The team optimized each base learner and then combined their votes using voting theory, testing both hard voting, where each classifier casts a single categorical vote, and soft voting, where the classifiers’ predicted class probabilities are averaged before a decision is made.

The results were striking. The best configuration, dubbed ADASYN-voting-soft, or AVS, achieved a mean accuracy of 90.25 percent across the reservoir classes. That figure represents an improvement of 12.22 to 16.84 percentage points over standalone machine learning methods. Just as important, the soft-voting ensemble proved more stable and more accurate than hard voting, especially for reservoirs whose well log signatures are ambiguous. This makes intuitive sense: when a rock sample sits near the boundary between two classes, individual classifiers may disagree, and hard voting can flip-flop from one measurement to the next. Averaging probabilities dampens that noise, producing smoother, more geologically plausible classifications along the borehole.

The AVS method also addresses a subtle risk in machine learning practice: the exclusion of small, high-quality sample sets. In geoscience, curated expert-labeled examples are expensive and scarce. A single classifier can effectively discard such precious minority data by underfitting them, but the ensemble approach compensates for the uncertainty of any individual model. Combined with ADASYN’s synthetic augmentation, the framework ensures that even thinly represented reservoir types exert real influence on the final prediction. The interpretability emphasis of the study matters here too, because geologists need to understand which logging features drive a classification before they commit drilling capital on the model’s advice.

Applied across the Xujiahe Formation’s second member, the workflow revealed the spatial distribution of reservoir types in the Xinchang area of the western Sichuan Depression, mapping where matrix-dominated zones give way to fracture-enhanced sweet spots. That spatial picture is more than an academic exercise. Accurate identification of high-quality reservoir intervals directly guides well placement, completion design, and production optimization in a basin that Chinese researchers have described as a super gas basin. Fracture-developed tight sandstone intervals hold disproportionately large shares of recoverable gas, so a tool that reliably flags them in uncored wells translates into fewer dry or underperforming wells.

The study also fits into a broader trend sweeping the geosciences: the migration of ensemble learning and imbalance-aware sampling from computer science into subsurface characterization. The authors’ bibliography traces a lineage from classic oversampling methods like SMOTE and Borderline-SMOTE through ADASYN’s adaptive variant, alongside a growing body of work applying random forests, boosting, and stacking to lithofacies identification, fracture detection, and fluid typing in basins from the Ordos to the Tarim. What distinguishes this contribution is the combination of a geologically meaningful classification scheme, an imbalance-correction step tuned to decision boundaries, and a probability-averaged ensemble, all applied to the notoriously tricky dual-medium problem.

Funded by China’s National Science and Technology Major Project and the National Natural Science Foundation of China, the work offers a template that could travel well beyond the Sichuan Basin. Any fractured tight reservoir, whether sandstone, carbonate, or volcanic, poses the same pairing of continuous well log data with rare, high-value target classes. As exploration pushes into ever deeper and more heterogeneous plays worldwide, tools like the AVS model suggest that the winning approach will not be any single clever algorithm, but disciplined engineering: classify the rocks in a way geologists understand, fix the data imbalance before it quietly biases the model, and let a well-calibrated team of classifiers vote where no individual model should be trusted alone.

Subject of Research: Machine learning classification of dual-medium tight sandstone gas reservoirs in the Sichuan Basin

Article Title: Interpretable Prediction of Dual-Medium Tight Sandstone Reservoir Types of the Sichuan Basin: A Multi-ensemble Learning Approach for Imbalanced Data Optimization

Article References: Zhang, J., Yue, D., Wang, W., Qu, L., Liu, Y., Li, W., Jin, W., & Fu, J. (2026). Interpretable Prediction of Dual-Medium Tight Sandstone Reservoir Types of the Sichuan Basin: A Multi-ensemble Learning Approach for Imbalanced Data Optimization. Natural Resources Research. https://doi.org/10.1007/s11053-026-10761-5

Image Credits: AI Generated

DOI: 10.1007/s11053-026-10761-5

Keywords: tight sandstone, Sichuan Basin, reservoir classification, machine learning, ensemble learning, ADASYN, imbalanced data, well logging, natural fractures, Xujiahe Formation, natural gas, dual-medium reservoir