AI Reads Well Logs to Map Hidden Volcanic Oil Reservoirs With 98.79% Accuracy

Deep beneath the Junggar Basin in northwestern China, layers of Carboniferous volcanic rock hold some of the country’s most promising oil and gas reservoirs. But telling one type of volcanic rock from another using only the electrical and physical measurements recorded by downhole instruments has long been one of the trickiest problems in petroleum geology. A team of researchers led by Bowen Xue and Zhaohui Zhang of Xinjiang University, working with colleagues from CNPC Bohai Drilling Engineering, the Xinjiang Branch of the Research Institute of Petroleum Exploration and Development, the University of Miami, and China National Logging Corporation, now reports a solution that borrows a page from modern computer vision. Their method, published in Earth Science Informatics, converts ordinary well logs into images and lets an attention-enhanced deep neural network read them, achieving a validation accuracy of 98.79 percent in identifying volcanic lithologies in the Dixi area of the Junggar Basin.

Volcanic rock reservoirs are increasingly important targets for global oil and gas exploration, but they are notoriously difficult to characterize. Unlike sedimentary strata, volcanic sequences display a wide variety of lithologies, are built up through multistage vertical superimposition of successive eruptions, and can change facies rapidly over short distances. All of this produces complex and variable logging-response characteristics, which makes lithology identification a major technical bottleneck in this type of exploration. In the Dixi area specifically, the Carboniferous system comprises a patchwork of basalt, andesite, dacite, rhyolite, volcanic breccia, and tuff, each with different porosity and fracture patterns that control where hydrocarbons can accumulate. Misclassifying a lithology in a well can therefore mean misjudging an entire reservoir.

The conventional approach relies on crossplots of logging values and hand-drawn response envelopes, methods that struggle when several lithologies overlap in the same measurement space. Machine learning classifiers such as support vector machines, random forests, AdaBoost, and XGBoost have improved matters, but they treat each depth point as an isolated set of numbers, ignoring the vertical context in which the measurements sit. The new study takes a fundamentally different route: it treats the logging curves themselves as images, so that a convolutional neural network can learn both the values of the curves and their shapes and spatial relationships, much as it would learn to recognize objects in photographs.

The workflow begins with rigorous geological ground truth. The researchers identified and calibrated the target lithologies using core thin sections, the microscopic slices of rock that provide the definitive record of what is actually present at a given depth. On that calibrated basis, they selected five logging curves as lithology indicator factors: natural gamma ray (GR), density (DEN), acoustic transit time (AC), compensated neutron log (CNL), and resistivity (RT). These five measurements capture complementary aspects of the rocks, from radioactive mineral content to bulk density, porosity, and fluid-filled fracture networks, and together they form the raw signal from which the lithologies must be distinguished.

Next comes the image conversion, the step that sets this method apart. The logging responses of the various lithologies are transformed into image samples, effectively rendering the five curves as visual patterns in which the morphology of each curve, its peaks, troughs, and transitions, becomes a spatial feature that a vision network can process. The researchers then extract features from these images using the ResNet18 architecture, a widely used 18-layer residual convolutional neural network. Crucially, they drive ResNet18 with transfer learning, meaning the network starts from weights already trained on a large general image dataset and is fine-tuned on the logging images. This strategy allows the model to reach high performance without needing an enormous library of labeled volcanic rock samples, a persistent constraint in subsurface geology where core data are expensive and scarce.

The real innovation, however, lies in the attention mechanism. The team introduced a convolutional block attention module, or CBAM, into the ResNet18 framework, creating a hybrid ResNet18-CBAM classifier. CBAM works by sequentially applying two forms of attention to the feature maps flowing through the network. A channel attention module learns which feature channels, which abstract pattern detectors, are most informative for the classification task, while a spatial attention module learns which regions of the image deserve the most weight. In practical terms, the network is forced to optimize its feature selection mechanism, focusing on the parts of the logging-image pattern that genuinely discriminate between, say, a tuff and a volcanic breccia, and suppressing irrelevant background variation. Attention modules of this kind have proven their worth across engineering applications, from pipeline defect diagnosis to flow pattern recognition, and this study demonstrates their value for subsurface lithology work as well.

When the researchers put the ResNet18-CBAM model to the test on validation samples from the Dixi area, it achieved an accuracy of 98.79 percent in volcanic lithology identification. That figure represents a level of reliability that hand-calibrated crossplot methods and earlier machine learning classifiers have struggled to approach in this geologically messy setting, where rapid facies changes and multistage superimposition blur the boundaries between rock types. The attention-enhanced network, in effect, learned to read the logging curves the way an experienced petrophysicist reads them, with an eye for the telling details, but with a consistency and speed no human analyst can match across thousands of meters of borehole.

The implications reach well beyond one basin. The authors argue that the proposed method provides an effective and reliable solution for volcanic lithology classification and contributes to the fine-scale characterization and evaluation of volcanic reservoirs. Fine-scale lithology maps are the foundation upon which everything else in volcanic reservoir evaluation rests: fracture distribution models, porosity estimates, weathered crust reservoir models, and ultimately production capacity predictions. Previous work in the same region has shown that quantitative characterization of fracture-pore distribution in the weathered volcanic crust of the Dixi area directly controls well productivity, and that electrical imaging logging can quantify volcanic fracture distribution. A fast, accurate, automated lithology classifier feeds directly into that chain of models, turning raw well logs into the geological input those models require.

The study also fits into a broader movement sweeping through petroleum geoscience, in which deep learning methods are being applied to nearly every stage of subsurface interpretation. Recent literature documents convolutional neural networks classifying lithology from drill core images, CNNs with sliding windows recognizing lithology and predicting porosity from well logs, dual attention-based models identifying lithology while drilling, geological information-driven deep learning for lithology identification, and hybrid CNN-SVM methods for shale reservoirs. The present study adds a distinctive contribution to this landscape by combining logging-curve morphology features, transfer learning, and a dual attention module in a single workflow tailored to the specific challenges of volcanic sequences, where the sheer diversity of rock types has always outpaced the discriminating power of simpler classifiers.

There are practical constraints to keep in view. The well-logging data and core data used in the study were provided by PetroChina Xinjiang Oilfield Company and remain subject to ownership and confidentiality restrictions, so the authors cannot publicly release the original dataset. Reproducing the 98.79 percent result elsewhere will therefore require partners with access to comparable calibrated core and log data. Still, the architecture itself, ResNet18 enhanced with CBAM and driven by transfer learning, is openly described and readily adaptable, and the underlying idea, that logging curves can be read as images by attention-equipped vision networks, is one that any exploration team with calibrated wells can pursue. As volcanic reservoirs assume a growing share of global exploration portfolios, from the Junggar Basin to the Songliao Basin and beyond, tools of this kind may well become as standard in the petrophysicist’s kit as the crossplot once was, quietly transforming how the industry sees the rocks at the bottom of the well.

Subject of Research: Machine learning identification of volcanic lithology from well logs in the Junggar Basin

Article Title: Volcanic lithology identification from well logs using a CBAM-enhanced CNN: a case study of the carboniferous system in the Dixi area, Junggar Basin

Article References: Xue, B., Zhang, Z., Fang, T., Li, Y., Jiang, N., & Ma, Y. (2026). Volcanic lithology identification from well logs using a CBAM-enhanced CNN: a case study of the carboniferous system in the Dixi area, Junggar Basin. Earth Science Informatics, 19(12), Article 217. https://doi.org/10.1007/s12145-026-02275-8

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02275-8

Keywords: volcanic lithology, well logs, convolutional neural network, CBAM, ResNet18, transfer learning, Junggar Basin, Dixi area, Carboniferous, oil and gas exploration, volcanic reservoirs, image recognition