Autonomous inverse modeling of complex groundwater systems via a physics-integrated large language model multi-agent...

Area of Science:

• Environmental Science
• Geosciences
• Artificial Intelligence

Background:

• Inverse modeling in subsurface hydrology faces challenges translating conceptual models to numerical simulations.
• Data-driven machine learning models often lack mass conservation and interpretability for regulatory use.

Purpose of the Study:

• To introduce Hydro-Agent, a physics-integrated multi-agent framework using LLMs to automate rigorous process-based simulations.
• To overcome the conceptual-numerical gap and operational bottlenecks in groundwater modeling.

Main Methods:

• A multi-agent framework coupling a reasoning agent (HydroCoder) and an execution agent (Executor).
• Utilizes a "code-as-policy" paradigm with LLMs to govern simulations and optimize solver strategies.
• Validated across heterogeneous media, kinetic experiments, and a field-scale aquifer.

Main Results:

• Successfully bridged the conceptual-numerical gap in groundwater flow and transport modeling.
• Autonomously identified geological features, achieved high-fidelity parameter retrieval under uncertainty, and reproduced complex field-scale trends.
• Maintained numerical rigor while enhancing operational autonomy and ensuring geologically plausible outcomes.

Conclusions:

• Hydro-Agent provides a transparent, auditable, and physically constrained AI modeling framework.
• Broadens access for hydrogeologists by abstracting computational complexity, prioritizing insights over implementation.
• Facilitates objective-driven modeling workflows in subsurface hydrology.