LMAFNet: Lightweight multi-scale adaptive fusion network with vertical reservoir information for lithology identification

Abstract

Lithology identification is fundamental to stratigraphic evaluation and reservoir characterization, which is crucial for assessing the potential value of oil and gas resources. However, traditional models often demonstrate limited accuracy due to the replication of non-linear relationships and class imbalance issues in well log data. In addressing these challenges, a lightweight multi-scale fusion network (LMAFNet) was proposed to mitigate data imbalance bias and improve lithology identification accuracy. The model integrates lithology vertical reservoir information using multi-scale preprocessing and, through its Multi-Scale Adaptive Module (MSAM), autonomously adapts neuron proportions for varying receptive fields based on lithology data size. Furthermore, the model performance is significantly improved by employing the channel attention mechanism and the focal loss function. To validate the efficacy of our proposed approach, extensive experiments were performed on the Daqing and Xinjiang datasets in China. Experimental results from the Daqing and Xinjiang datasets underscore the model’s superior performance, with accuracies reaching 94.25% and 95.91%, respectively, and marked enhancements in recall and F1 scores. Additionally, the method’s practical application in blind well lithology identification on the Daqing dataset has been validated, demonstrating its adaptability and utility across different production environments. This method merits application and promotion in the domain of lithologic reservoir identification.