AutoSurrogate: An LLM-Driven Multi-Agent Framework for Autonomous Construction of Deep Learning Surrogate Models in Subsurface Flow

Researchers Jiale Liu and Nanzhe Wang have introduced AutoSurrogate, a novel framework that leverages large language models (LLMs) to orchestrate a team of four specialized AI agents. These agents collaborate to autonomously construct deep learning (DL) surrogate models for computationally intensive subsurface flow simulations, a critical task in fields like geological carbon storage. The system is designed to bridge a significant expertise gap, allowing domain scientists—who may lack deep machine learning knowledge—to build high-quality models through simple natural-language instructions. Given simulation data and optional user preferences, the framework can produce a deployment-ready model with minimal human intervention, handling complex steps from data profiling to final quality assessment.

In a practical demonstration, AutoSurrogate was applied to a 3D geological carbon storage modeling task, which involved mapping permeability fields to pressure and CO2 saturation over 31 timesteps. The system autonomously executed a full pipeline: profiling the input data, selecting an architecture from a pre-defined model zoo, performing Bayesian hyperparameter optimization, training the model, and rigorously assessing its quality against user-specified thresholds. Crucially, it also managed common failure modes, such as restarting training with adjusted configurations when numerical instabilities arose or switching architectures if predictive accuracy was insufficient. Without any manual tuning, AutoSurrogate outperformed both expert-designed baseline models and general-purpose AutoML methods, showcasing its strong potential for real-world, complex scientific and engineering applications where simulation speed is paramount.