STNet: Advancing Lithology Identification with a Spatiotemporal Deep Learning Framework for Well Logging Data

Abstract

In the realm of oil and gas exploration, accurate identification of lithology is imperative for the assessment of resources and the refinement of extraction strategies. While artificial intelligence techniques have garnered considerable success in lithology identification, existing methodologies encounter difficulties when addressing highly heterogeneous and geologically intricate unconventional oil and gas reservoirs. Specifically, they struggle to account for the dynamic variations in sample characteristics across spatial dimensions and temporal sequences. This separate treatment of spatial and temporal dynamics not only confines the precision of fluid prediction but also significantly attenuates the robustness of the models. To address this challenge, we propose the spatiotemporal network (STNet), a dual-branch deep learning framework that integrates seamlessly spatial feature graph methods with time-sequential prediction methods. By employing a graph structure that accounts for spatial characteristics to capture the complex spatial relationships within logging data, and by utilizing a temporal model to discern the dynamic properties of time series data, this dual-mechanism framework enables a more comprehensive understanding of the multidimensional attributes of subsurface fluids, thereby enhancing the accuracy of lithology identification. Experimental results from multiple wells in different regions of the Tarim and Daqing oilfields demonstrate that STNet not only achieves detection accuracy exceeding 95% but also exhibits strong generalizability. The results indicate a significant improvement in the accuracy of lithology identification compared to seven other advanced models. Integrating both temporal and spatial elements of logging data provides a new perspective for enhancing fluid prediction capabilities.