Abstract We introduce a novel large-language-model (LLM) driven agentic framework to automate end-to-end surrogate modeling of simulator-driven workflows in the oil and gas domain. The autonomous agent orchestrates critical steps—initial sampling, simulator execution, adaptive retraining, and strategy switching—to minimize expensive simulator calls while meeting accuracy targets with minimal subject matter expert (SME) intervention. Applied to well-network flow and gas processing simulations, the agent dynamically switches sampling strategies to achieve user-specified accuracy efficiently. In each iteration, the LLM agent generates candidate input sets, evaluates them via the simulator, and fine-tunes the surrogate model on the expanded dataset. Uncertainty estimates from Monte Carlo (MC) dropout guide an acquisition function that scores candidates (e.g., by residual error or predicted variance) to prioritize informative samples. The agent monitors the surrogate model’s error on a validation set and triggers early stopping to prevent overfitting. Initial results demonstrate that the agent-driven approach can autonomously produce high-fidelity surrogates with intelligent sampling and model updates. The adaptive strategy switching yielded roughly a 6% lower Symmetric Mean Absolute Percentage Error (SMAPE) versus using any single sampling method, and the pipeline converged to the target error threshold with minimal human guidance. This framework is among the first to employ an LLM as an autonomous orchestrator for surrogate modeling workflows, significantly reducing computational cost and SME effort while providing a plug-and-play solution for complex simulator-driven engineering problems.
Shetty et al. (Mon,) studied this question.