AutoSurrogate: An LLM-Driven Multi-Agent Framework for Autonomous Construction of Deep Learning Surrogate Models in Subsurface Flow
AutoSurrogate is an LLM-driven framework that automates the construction of deep learning surrogate models for subsurface flow simulation, enabling domain scientists without machine learning expertise to build high-quality models through natural language instructions. The system autonomously handles data profiling, architecture selection, hyperparameter optimization, and quality assessment while managing failure modes, demonstrating superior performance to expert-designed baselines on geological carbon storage tasks.
AutoSurrogate addresses a critical bottleneck in scientific computing: the expertise gap preventing domain scientists from leveraging deep learning acceleration. Subsurface flow simulation is computationally expensive for uncertainty quantification and data assimilation tasks, making surrogate models valuable. However, constructing effective deep learning surrogates requires specialized knowledge in neural network architecture design, hyperparameter tuning, and model validation that most geoscientists lack. This expertise barrier has limited broader adoption of surrogate techniques across scientific domains.
The framework represents a convergence of large language models with scientific computing automation. By deploying four specialized agents that collaborate on distinct tasks—data profiling, architecture selection, Bayesian optimization, training, and quality assessment—AutoSurrogate eliminates manual intervention at intermediate stages. The system autonomously recovers from common failure modes like numerical instabilities and insufficient accuracy, demonstrating sophisticated error handling. The ability to generate deployment-ready models from single natural-language instructions suggests a significant shift in how scientific computing workflows may be designed.
The demonstrated performance on 3D geological carbon storage modeling is particularly relevant given climate-focused computational needs. Outperforming both expert-designed baselines and domain-agnostic AutoML methods indicates the framework leverages domain-specific context effectively. This has broader implications for scientific research acceleration across geoscience, physics simulation, and engineering domains where high-fidelity simulation costs remain prohibitive. Organizations investing in scientific computing infrastructure should monitor this approach's maturation and potential integration into existing workflows.
- →AutoSurrogate enables non-ML experts to build deep learning surrogates through natural language, lowering the expertise barrier in scientific computing
- →The multi-agent framework autonomously handles architecture selection, hyperparameter optimization, and failure recovery without manual intervention
- →System outperformed expert-designed baselines on 3D carbon storage modeling, suggesting practical deployment readiness
- →Addresses computational bottleneck in uncertainty quantification and data assimilation for subsurface flow problems
- →Framework combines LLM capabilities with scientific domain knowledge for automated model construction workflows