Autoregressive model with discrete feature representation for well log interpretation

Accurate well log interpretation is essential for subsurface reservoir characterization, as it provides key information on lithology, stratigraphic structures, and fluid distributions. However, the complexity of well log data, limited labeled datasets, and insufficient exploitation of multimodal information remain major obstacles to reliable interpretation. Existing machine learning methods often rely on task-specific training and lack generalization across different interpretation objectives, restricting their applicability in large-scale geophysical analysis. To address these challenges, we propose a novel framework with discrete feature representation for well log interpretation. This framework integrates vector quantization with wavelet transformation to generate noise-resistant discrete representations, which are then processed by an autoregressive model via a next-token prediction structure. After discretization and standardization, multi-source well logs are fused into a semantically aligned discrete token sequence, facilitating the mining of cross-modal geological correlations. We further design a two-stage training strategy: unsupervised pre-training using random token permutations, followed by supervised fine-tuning guided by prompt-based tasks. Leveraging autoregressive training, the model can be flexibly adapted to specific interpretation tasks by modifying prompt tokens during the supervised fine-tuning stage. Extensive experiments on classification and regression benchmarks in the Ordos Basin demonstrate that the proposed method captures intrinsic correlations among well log curves, thereby improving the accuracy and robustness of reservoir characterization. • A novel discrete feature representation framework that enhances noise robustness and interpretability in well log analysis. • A tokenization strategy that integrates multi-source well logs into coherent sequences for effective cross-modal fusion. • An adaptive autoregressive pre-training and prompt-guided fine-tuning paradigm for well log interpretation tasks.