LLM and Digital Twin–Aided Decision Making for Water Governance

Problem: Groundwater governance is hindered by the high computational cost and latency of simulation models such as MODFLOW. Policymakers typically rely on manual, trial-and-error exploration of parameter spaces, which is inefficient and time-consuming. Additionally, the lack of intuitive interfaces and the need to manually retrieve regulatory information further slow down decision-making processes. 

Approach: We developed an integrated decision-support system combining a digital twin, optimization, and LLM-based assistance: 

• Digital Twin Interface: Built an interactive UI for configuring, executing, and comparing groundwater simulations. The system supports structured scenario input and visual comparison of outputs for rapid analysis.
• Bayesian Optimization: Incorporated to efficiently search the parameter space and identify optimal configurations, reducing reliance on exhaustive trial-and-error methods.
• RAG + LLM Integration: Implemented a Retrieval-Augmented Generation pipeline to provide real-time, context-aware answers to regulatory and policy-related queries, eliminating the need for external searches. 

Results: 

• Enabled seamless execution and comparison of MODFLOW simulations via a user-friendly interface
• Reduced decision-making time by automating parameter search using Bayesian optimization
• Improved accessibility by allowing natural language inputs to auto-populate simulation parameters
• Accelerated policy analysis through integrated regulatory Q&A using RAG-based LLMs 

Conclusion: This work demonstrates that integrating digital twins with optimization and LLM-based knowledge systems can significantly enhance groundwater governance. The framework improves efficiency, usability, and decision quality, making it highly suitable for policy-driven environmental management applications.

Committee: Simanta Mitra (major professor)