A multi-agent collaborative framework with human-in-the-loop for well log interpretation and applications

This paper proposes a multi-agent system centered on large language models to address the issues that traditional well log interpretation relies on expert experience and poses great difficulty in multi-disciplinary collaboration and constructs a digital twin architecture across three dimensions of agents, tools and environment. At the agent level, a role-based architecture is established to decompose the complex log interpretation workflow into independent subtasks, enabling structured transfer of expert knowledge. At the tool level, petrophysical formulas and machine learning algorithms are encapsulated to form a physics-data dual-path hybrid reasoning mechanism; at the environment level, a standardized digital twin space is established based on the Model Context Protocol to achieve closed-loop control of the entire workflow. Engineers can drive the system through natural language commands to complete the full log interpretation process from data loading and parameter calculation to reservoir classification, realizing end-to-end automation from raw data to interpretation conclusions. In tests on 100 field wells, the system generates key interpretation parameters that are highly consistent with expert results, exhibiting stable recognition capability for complex reservoir types. This study demonstrates that this human-machine collaborative working mode significantly enhances the standardization and efficiency of well log interpretation, providing technical reference for intelligent transformation of highly specialized industrial processes.