Log-driven classification of shale lithofacies employing Kolmogorov-Arnold network assisted by conditional generative adversarial network and transfer learning

Lithofacies classification is essential for stratigraphic interpretation and sedimentological characterization, but remains challenging in shale reservoirs because rapid lithological variations weaken the discriminative capability of conventional well logs. To improve shale lithofacies identification, this study proposes a hybrid classification framework, termed NCA-Bayes-KAN, in which the Kolmogorov-Arnold network (KAN) is used as the core classifier, neighbourhood component analysis (NCA) is employed for feature optimization, and Bayesian optimization (Bayes) is applied for hyper-parameter tuning. Geological and well-logging data from the Ordos Basin were used to evaluate the proposed framework. Five commonly used classifiers, including k-nearest neighbors, support vector machine, random forest, deep neural network, and extreme gradient boosting, were introduced as benchmark models. Experimental results based on accuracy, F1-score, Cohen’s kappa, and area under the curve (AUC) show that NCA-Bayes-KAN consistently outperforms the competing models. Because softmax-based outputs may produce poorly calibrated probabilities, probability calibration was further incorporated to improve the reliability of predicted class probabilities, leading to additional performance improvements. To enhance model generalization, conditional generative adversarial network (CGAN) and transfer learning were integrated into the framework. CGAN was used for data augmentation, while transfer learning enabled the use of pre-learned knowledge from the source well group. The results indicate that NCA-Bayes-KAN combined with probability calibration provides an accurate and reliable solution for shale lithofacies classification within the investigated study area, while CGAN and transfer learning further improve its transferability among the studied wells and well zones.