AutoSAM: an Agentic Framework for Automating Input File Generation for the SAM Code with Multi-Modal Retrieval-Augmented Generation¶

🕒 Published (v1): 2026-03-25 18:59 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

AutoSAM is a ReAct-based LLM agent framework that automates the generation of input files for the SAM nuclear thermal-hydraulics code by combining system instructions, multi-modal RAG over solver documentation, and seven specialized tools for parsing PDFs, images, and spreadsheets. The agent extracts simulation parameters from heterogeneous engineering documents into a human-auditable intermediate YAML representation, then synthesizes validated, solver-compatible input decks. Evaluated on four case studies of increasing complexity, it produces runnable SAM models with 100% structured-input utilization and ~88% PDF text extraction.

Problem¶

Constructing valid input files for system-level thermal-hydraulics codes like SAM is a labor-intensive, error-prone bottleneck requiring analysts to manually extract parameters from heterogeneous engineering artifacts (PDFs, P&IDs, spreadsheets, legacy reports) and transcribe them into solver-specific syntax—a process that can consume days to weeks per model. SAM is closed-source and absent from LLM training corpora, making direct LLM application unreliable. Prior automation approaches (OCR, rule-based extraction, GUIs) fail when critical information is fragmented across multi-modal sources or encoded visually in schematics.

Method¶

AutoSAM uses a three-pillar specialization strategy inside a LangChain-orchestrated ReAct agent:

System instructions: Expert-engineered persistent context encoding SAM's input block structure, modeling conventions, and few-shot examples.
RAG: SAM's user guide and theory manual embedded into a vector store (static directory); task-specific user documents embedded into a dynamic directory. Both are queried via semantic similarity search at inference time.
Seven specialized tools divided into:
Content parsing: Excel reader (tabular design data), text file reader.
Retrieval: PDF analysis tool (Nougat OCR → vision-LLM figure captioning → chunk embedding → similarity retrieval); image analysis tool (VLM description → embedding → retrieval).
Instruction passing: Input file creator tool (converts intermediate YAML → SAM input deck via structured LLM prompt); input file validator tool (deterministic syntax + physics checks: block completeness, units, geometry connectivity, boundary conditions).
Python code execution: general computation.

The document processing workflow is: ingest unstructured docs → extract via tools → produce intermediate YAML (human-in-the-loop checkpoint) → merge with structured inputs → generate SAM input deck → validate.

Key Contributions¶

End-to-end agentic harness for SAM input generation from raw engineering documentation, including multi-modal sources.
Multi-modal RAG pipeline integrating Nougat-based scientific text extraction, VLM figure interpretation, semantic embedding, and hybrid static/dynamic retrieval directories.
Intermediate YAML representation as a deliberate human-in-the-loop checkpoint between document extraction and final solver syntax, aligned with nuclear QA expectations.
Deterministic input validator combining syntax checks and physics verification (units, geometry connectivity, boundary/initial conditions).
Four progressive case studies covering single-pipe, multi-physics coupled, parallel-channel core (ABTR), and full primary-loop reconstruction (MSRE).

Results¶

100% utilization of structured (Excel/spreadsheet) inputs across all cases.
~88% parameter extraction rate from PDF text sources.
100% completeness in vision-based geometric extraction from figures/schematics.
All four case studies produced runnable SAM models whose outputs are consistent with expected thermal-hydraulic behavior.
Agent explicitly flags missing data and labels assumed values, preserving traceability.
Cases: (i) single-pipe steady-state, (ii) solid-fuel channel with temperature-reactivity feedback, (iii) ABTR reactor core (multi-modal, parallel-channel), (iv) MSRE primary loop (full-loop topology reconstruction from heterogeneous sources).

Limitations¶

Nougat model cannot extract information embedded inside PDF figures; a separate VLM pass is required, adding pipeline complexity and potential description quality variance.
System instructions are constrained by the LLM context window; only the most essential SAM conventions are included, and performance degrades for constructs not covered.
The agent may hallucinate or apply incorrect engineering judgment when data is absent; assumed values are labeled but not independently verified.
Evaluation is limited to four SAM-specific cases; generalizability to other system codes (RELAP, TRACE, CATHARE) is not demonstrated.
No quantitative error analysis on the physics fidelity of generated files beyond "consistent with expected behavior"—no reference simulation comparisons with numerical metrics are reported in the provided text.
Framework depends on a closed-source, licensed code (SAM), limiting reproducibility.

Relevance to Harnesses / Meta-Harnesses¶

AutoSAM is a domain-specific simulation harness: it wraps a ReAct LLM agent with a structured tool registry, a dual-mode RAG knowledge layer, and a deterministic validation stage to automate a complex, multi-step engineering workflow end-to-end. The intermediate YAML checkpoint is a canonical harness pattern—separating extraction from generation, enabling human intervention, and enforcing a typed intermediate representation before downstream code synthesis. The static/dynamic RAG directory split mirrors how meta-harnesses manage reusable vs. task-specific knowledge. AutoSAM directly extends the AutoFLUKA precedent (Monte Carlo input generation) and illustrates a replicable harness template: (system instructions + RAG + tool suite + intermediate IR + validator) applicable to any solver-specific automation task.