Scaling the Expert Eye: Automating Subsurface Model Evaluation and Tuning via Reservoir Engineering Inspired AI Agents

Abstract

Abstract Predicting the behavior of oil fields is a notoriously complex endeavor, often relying on high-fidelity predictive models (e.g. simulation models) that are sensitive to a myriad of geological and operational parameters. Traditionally, the "gold standard" for refining these models has been the manual intervention by reservoir engineers. These experts possess the nuanced domain knowledge required to interpret data anomalies and perform "history matching"—adjusting model inputs to align with historical performance. However, this human-centric approach is often a bottleneck; it is slow, difficult to scale, and limited by the speed at which a human can iterate through high-dimensional data sets. To address this, we introduce an autonomous AI agent designed to bridge the gap between expert intuition and computational efficiency. In order to perform this task autonomously, the AI Agent needs the ability to do the following: (1) Validate a model's quality using the same qualitative and quantitative metrics as used by the experts (2) Determine what needs to be changed and how to change them to improve the model (3) Modify these parameters and evaluate the model in order to improve the model (4) Repeat the process until all metrics are satisfied. This is achieved by using state-of-the-art LLMs (large language models) with system prompts fed with documentation of all the steps above, along with LangChain-based agent architecture developed specifically for this purpose. By codifying the analytical workflows and heuristic strategies used by veteran reservoir engineers, the agent performs automated parameter tuning across predictive models. The system leverages an assisted data assimilation approach already in place, using rigorous back-testing to evaluate the impact of each adjustment in real-time. Instead of a brute-force search, the agent applies specialized logic to identify which "knobs" to turn—such as permeability or porosity ranges —to generate results that best mirror reality, effectively mimicking the "expert eye" at scale. The integration of this agent transforms the modeling pipeline from a labor-intensive manual task into a high-speed, expert-driven optimization process. Initial results tested on multiple real field models demonstrate that the agent can achieve model fidelity comparable to manual tuning while significantly reducing the time-to-insight from weeks to hours, and also minimizing critical errors that might be introduced by human engineers due to the complex model validation process. By automating the iterative back-testing/history-matching cycle, organizations can democratize the specialized knowledge of their top engineers across entire portfolios of assets. This ensures that predictive models remain robust, reliable, and grounded in the practical realities of field operations, allowing human engineers to focus on high-level strategic decision-making rather than repetitive parameter tweaking.