Abstract: To address the challenges of complex nonlinear log data processing and low accuracy in lithology identification, we propose a novel approach for carbonate identification, which combines an improved log-centered principal component analysis (LC-PCA) with a back propagation neural network (BPNN) optimized by reverse-driven Cuckoo Search (RDCS). The LC-PCA method effectively captures the nonlinear features of log curves and achieves dimension reduction through logarithmic transformation and data centralization. To overcome the limitations of the traditional Cuckoo Search algorithm, a dimension-wise reverse learning strategy and the rider optimization algorithm (ROA) are incorporated to establish the RDCS algorithm, which significantly enhances global search and convergence of the traditional algorithm and achieves optimal performance in solving benchmark functions, e.g. Sphere and Rosenbrock. The RDCS-BPNN model for lithology identification is finally constructed by utilizing RDCS to optimize BPNN weights and thresholds. As per a case study of the carbonate gas reservoirs of the Council Grove Formation in Panoma field, the RDCS-BPNN model achieves outstanding performance in discriminating 9 lithology types, with an average accuracy of 93.64% and an area under the curve (AUC) of 0.926 based on the micro-average receiver operating characteristic (Micro-ROC) curve for model evaluation. Compared with mainstream approaches, the RDCS-BPNN model achieves an accuracy improvement of 3.5%–10.3% and demonstrates excellent capability in identifying the transition zones between dolomites and wackestones. The RDCS-BPNN predictions exhibit high agreement with actual lithologies in the complex intervals of experimental wells. Therefore, the integrated framework of LC-PCA for data processing and RDCS-BPNN offers a reliable technical solution for lithology, pore structure, and heterogeneity characterization in complex carbonate reservoirs.