State Machine Structured Agents for Physical Science Reasoning

Abstract

Large language models (LLMs) have shown promise as scientific assistants capable of reasoning, tool invocation, and autonomous analysis. However, their use in the physical sciences remains limited by the need for strong guarantees of explainability, traceable reasoning over long computational workflows, and the high consequences of subtle errors. Many existing LLM-based agent systems rely on unstructured conversational context or loosely coupled tool calls, which are insufficient for scientific domains governed by strict physical constraints and nonlinear dynamics. In this study, we explore the use of an AI assistant for an Earth system science use case. Unique to this work, we introduce a state machine agentic architecture that enables LLMs to perform structured, physically grounded scientific analysis through an explicit and enforceable notion of agent state. We define the agent state as a structured tuple, maintained across agent iterations using explicit update rules and strongly typed interfaces that enforce physical feasibility during tool invocation. We evaluate this methodology for a case study of ocean overturning circulations and a potential slowing or collapse of this system. The AI assistant interacts with a reduced-order ocean model and in-house developed deep learning models. The AI assistant is able to decompose complex questions into actionable steps that are answered using the set of tools developed. Compared to traditional grid search approaches, the agent autonomously designs parameter explorations, invokes high-dimensional models within valid physical bounds, and produces interpretable scientific summaries while reducing simulation count. The results demonstrate that the structured agent state is a key enabler of reliable LLM-based scientific assistants and is broadly applicable to constrained tool-centric scientific workflows.