PlasmaAura: A Multimodal LLM-Based Agent for WEST Tokamak Data Analysis

Abstract

Tokamak operations generate large volumes of diagnostic data needed for plasma control and for analyzing complex plasma dynamics. During experimental sessions, session leaders and researchers must review the previous pulse and define the next experimental configuration within a 10–15-min interval, making rapid access to relevant information essential. Effective decision support often relies on dedicated tools for retrieving information from operational data systems, typically based on SQL databases or flat files. However, the manual query interfaces to these data sources often present a steep learning curve, thereby offering limited flexibility for session leaders and physics operators in the dynamic environment of tokamak operations. Recent advancements in artificial intelligence (AI), particularly in large language models (LLMs), offer a promising approach for interacting with scientific data through natural language. This work introduces PlasmaAura, a conversational agent based on a multimodal LLM (MLLM) informed with tokamak-specific knowledge through prompt engineering. This agent is capable of interpreting natural language queries to retrieve, correlate, and visualize both diagnostic and contextual tokamak data, producing executable code, figures, and textual explanations as verifiable outputs. In the present work, multimodality refers to the joint reasoning of the MLLM over textual context, generated code, and visual outputs, while the input data considered remain time-series diagnostics. We present case studies covering data retrieval, event detection, and descriptive analysis, with measured response times confirming suitability for interpulse use. This work represents a first step toward integrating AI agent tools into tokamak operation preparation, with the aim of enhancing decision-making workflows and supporting ITER objectives.