LLM-Orchestrated Digital Twins for Safe, Human-Centered Decision Support in Precision Agriculture

Charles Cao, Sajal Das, Jie Zhuang, Robert Davis

Keywords: digital twins, precision agriculture, large language models, decision support systems, safety constraints, natural language interfaces, agricultural AI, cyber-physical systems

TL;DR: LLM-orchestrated digital twins enable safe, explainable agricultural decision support by combining biophysical and operational models with natural language interfaces, achieving 87.5% intent grounding accuracy and 88% safety enforcement success.

Abstract: Digital technologies in agriculture (drones, IoT sensors, satellite imagery, and precision machinery) have created an opportunity for AI systems that not only forecast outcomes but also recommend and safely execute field-level interventions. We present a digital twin-based decision support framework in which a tool-using Large Language Model (LLM) agent operates over a dual twin of crop--soil dynamics and farm operations. Farmers express high-level goals and constraints in natural language (e.g., reduce fungicide use by 10% without risking major outbreaks), which are compiled into machine-readable Field Management Cards that encode multi-season objectives, safety constraints, and regulatory limits. The LLM agent plans candidate interventions (e.g., modified spray schedules, irrigation adjustments) by invoking twin services and farm asset APIs (sprayers, drones, sensors) through a standardized tool interface. Digital twins enforce feasibility and safety before recommendations are surfaced or actions are dispatched. Through comprehensive evaluation on 85 test cases, we demonstrate that the system achieves 87.5% accuracy in mapping natural language queries to appropriate tool calls and 88% success rate in enforcing safety constraints such as pre-harvest intervals and irrigation limits. Critically, the digital twin successfully repairs 100% of detected violations, ensuring fail-safe operation. Our results illustrate a path toward robust, human-centered AI assistants that connect agricultural data, domain models, and actuation in a single decision loop while maintaining safety and regulatory compliance.

Submission Number: 27