A Geology-Guided GBDT–BiSSM Framework for Carbonate Reservoir Classification from Well Logs

This is the authors' abstract. We don't add key points for this paper.

Accurate reservoir classification is essential for the efficient development of carbonate hydrocarbon reservoirs. Strong heterogeneity, complex diagenetic alteration, and overlapping well-log responses make this task difficult in the low-porosity and low-permeability Jialingjiang Formation of the central Sichuan Basin. To address this challenge, a GBDT-BiSSM fusion framework is proposed by integrating gradient boosting decision trees with a multiscale bidirectional state-space model. GBDT extracts static discriminative information from multivariate logging responses, whereas BiSSM learns depth-dependent contextual representations from ordered well-log sequences to characterize vertical variations and boundary-related transitions. Geological information is not embedded as an explicit constraint in the model-training process; instead, the fused predictions are refined during inference using geological-consistency decoding based on reservoir thickness, evaporite-related responses, and vertical reservoir continuity. Experimental results show that the proposed method achieves an overall classification accuracy of 92.9% and effectively reduces isolated label fluctuations and the misclassification of thin reservoirs. The method provides an effective approach for the detailed delineation of low-porosity and low-permeability carbonate reservoirs and the prediction of high-quality reservoir intervals.