Reservoir Modeling Assisted with Knowledge-Enhanced Large Language Models

Reservoir simulation is an important technique to characterize subsurface fluid flow and to support reservoir development and production optimization. Conventional reservoir modeling requires engineers to extract model-relevant information from heterogeneous data, organize the information into predefined formats, and generate simulator-specific input files through professional software. Constructing and iteratively updating such models can often take months of effort due to the high modeling complexity. Large language models (LLMs) offer a potential alternative by translating modeling requirements into simulator input instructions through natural language modeling instructions. However, proprietary reservoir simulators are poorly represented in public training corpora, and their keyword systems may differ substantially from those of widely documented commercial simulators. This study addresses these challenges using the Compositional Advanced Simulation System (COMPASS), an in-house reservoir simulator for unconventional resources, as the target simulator. The proposed knowledge-enhanced LLM framework combines retrieval-augmented generation (RAG) with schema-constrained tool calling to support well-control instruction generation. On the 50-task RECURRENT benchmark using DeepSeek-V4.1-Flash, the four generation strategies achieved average accuracies of 0.67%, 74%, 98%, and 100% for Strategy I (zero-shot), Strategy II (RAG), Strategy III (schema-constrained tool calling), and Strategy IV (RAG combined with schema-constrained tool calling), respectively. Additional experiments with different base LLMs showed accuracies ranging from 92.0% to 100.0%. These results demonstrate that combining enterprise-domain knowledge retrieval with schema-constrained output generation can substantially improve the reliability of LLM-based generation for proprietary reservoir simulation software.