End-to-End Depositional Microfacies Identification from Variable-Log Inputs Using CNN and Attention-CRF

Reliable identification of sedimentary microfacies from well logging data is essential for reconstructing depositional environments and conducting fine-scale reservoir characterization, yet remains challenging due to complex curve morphology, strong vertical heterogeneity, and limited logging information in mature fields. Although recent machine learning and deep learning methods have improved automation and overall accuracy compared with traditional manual interpretation, most existing approaches still suffer from strong subjectivity, inaccurate facies boundary prediction, and heavy dependence on multiple logging curves, which restrict their applicability and lead to degraded performance in data-scarce mature reservoirs. To explicitly address these challenges, this study proposes an end-to-end sedimentary microfacies identification framework that requires as few as a single logging curve and directly processes raw logging sequences of variable lengths, thereby avoiding curve distortion caused by stretching or artificial fixed-length segmentation. The proposed framework integrates convolutional neural networks, a Transformer-based self-attention mechanism, and a conditional random field to capture characteristic curve morphology, model long-range stratigraphic dependencies along the well section, and enforce facies continuity at the sequence level. Experimental results demonstrate strong agreement between the predicted sedimentary microfacies and core observations. In addition, visualization analyses of convolutional kernel responses and CRF-based optimization effects provide supporting evidence for the interpretability of the proposed framework, highlighting its practical applicability for sedimentary microfacies identification in mature fields with limited logging data.