Using Generative AI to Build a Reservoir Simulation Assistant

Abstract Numerical reservoir simulation is an intricate aspect of reservoir engineering, requiring a thorough understanding of reservoir engineering principles and the specific syntax of the reservoir simulation software used with the objective of designing an economical Field Development Plan (FDP) that is used to extract the hydrocarbons in real life which entails a considerable investment. This complexity is not unique to reservoir engineering but is also present in domains such as climate modeling, aerospace engineering, and electrical grid simulation, where accurate modeling and simulation are vital. Reservoir Engineers spend a considerable amount of time working with input "decks", which are structured text files used to build reservoir simulation models. Despite available tools to aid in their creation, manual refinement is often necessary to optimize simulation outcomes for forecasting or to align simulation results with historical data in history-matching processes. This task requires substantial expertise and can be time and effort consuming. Another challenge engineers often face is converting simulation models between different reservoir simulators. This translation process is complex due to variations in keywords, record formats, default values, and functionalities across different simulators, which may require substituting features with similar but not identical ones. Generative AI (Gen AI) holds significant promise in addressing these challenges by supporting the creation, modification, and translation of reservoir simulation models input decks thus eliminating repetitive tasks and thereby reducing the manual and tedious effort required, learning time and minimizing errors. Crucially, Gen AI can also play a transformative role in analyzing existing models and interpreting simulation results. This includes identifying potential issues within models, providing valuable insights during the Field Development Plan creation process, and supporting decision-making through in-depth analysis of simulation results. These capabilities are beneficial for a wide range of tasks, including history matching, uncertainty quantification, and forecasting. To achieve this, Gen AI should be integrated with specialized agents - programs or scripts that retrieve and provide the necessary contextual information and that possess limited analytical capabilities when scanning input and results/log files. This paper introduces the architecture of an intelligent reservoir simulation assistant aimed at increasing the productivity of reservoir engineers, reducing repetitive tasks, and supporting less experienced engineers in honing their skills. By leveraging the strengths of Gen AI supported by analytic agents, this assistant aspires to make reservoir simulation workflows more efficient, accurate, and conducive to informed decision-making in reservoir management. To the best of the authors’ knowledge, the use of Gen AI as a Reservoir Simulation Assistant is still an evolving topic in multiple technology and operator companies in the oil and gas industry, which ensures the novelty of this paper and the foreseen benefits of publishing it to the community.