AutoSAM automates SAM input-file creation with agentic multi‑modal retrieval-augmented generation

What the paper proposes

A new preprint on arXiv (arXiv:2603.24736) introduces AutoSAM, an "agentic" framework that aims to automate the labor‑intensive task of constructing input files for the System Analysis Module (SAM), a system‑level thermal‑hydraulics code used in the design and safety analysis of advanced nuclear reactors. The authors describe a pipeline that couples multi‑modal retrieval‑augmented generation (RAG) with agentic orchestration: the system extracts design parameters from heterogeneous engineering documents (text, tables, figures) and translates them into solver‑specific syntax for SAM. Reportedly, the approach reduces manual reconciliation of design data and speeds the generation of ready‑to‑run input decks.

Why this matters

For reactor engineers and safety analysts, input file preparation is a major bottleneck — a lot of domain knowledge lives scattered across CAD drawings, spec sheets, and legacy reports. AutoSAM promises to bridge those formats and cut turnaround time. For Western readers unfamiliar with China’s tech scene or arXiv’s role in open research, arXiv is a common repository for early‑stage AI and engineering work; this paper is presented as a new submission on that platform and can be accessed at https://arxiv.org/abs/2603.24736. The technical angle is straightforward: apply large‑language and multi‑modal models not just to summarize documents, but to execute discrete, rules‑driven transformations required by engineering solvers.

Risks, verification and policy questions

The gains are clear, but so are the red flags. Can an automated agent reliably capture implicit assumptions and safety margins embedded in engineering judgment? Who validates the machine‑generated inputs before they are used in safety analyses or licensing submissions? It has been reported that similar AI‑assisted tools in safety‑critical domains raise questions about traceability, auditability and regulatory acceptance. There is also a geopolitical dimension: tools that lower the barrier to nuclear‑system modeling could intersect with export controls and non‑proliferation policies, complicating cross‑border collaboration and deployment. AutoSAM is an intriguing technical step; adoption will hinge on rigorous validation, transparent provenance, and alignment with regulatory frameworks.