Speaker
Description
Monte Carlo particle transport codes are essential tools for accelerator shielding design, radiation safety assessment, and dose evaluation. However, their practical use requires detailed knowledge of input syntax, geometry modeling, source definition, tally selection, statistical uncertainty evaluation, and output interpretation. Recent progress in large language models and AI agents may provide a new way to support such complex workflows.
In this presentation, we introduce AI-assisted workflows developed for PHITS as a practical example. An AI-readable knowledge base was prepared from PHITS manuals, lecture materials, recommended inputs, sample files, utility information, and developer-curated cautions. This knowledge base was used for conversational PHITS support with NotebookLM and for AI-agent-driven execution and analysis workflows using agents such as Codex and Claude Code. Demonstration tasks showed that AI agents could edit PHITS input files, execute calculations, inspect errors, perform post-processing, and assist with shielding optimization when appropriate knowledge resources and operational rules were provided.
Based on these results, we would like to discuss at the meeting the future role of AI agents in shielding calculations, particularly their potential to execute multiple Monte Carlo codes under comparable conditions, support benchmarking and intercomparison, and help users focus on physical interpretation, uncertainty assessment, and engineering judgment.
| Topics | T2. Code status, development and model converters |
|---|