AI Takes the Reactor: Automating Complex Nuclear Simulations

by

Author: Denis Avetisyan

A new framework leverages artificial intelligence to drastically reduce the time and effort required to prepare input files for nuclear reactor modeling.

The system’s code input file structure, as detailed in Figure 2, establishes a foundational framework for the SAM system’s operational logic.
The system’s code input file structure, as detailed in Figure 2, establishes a foundational framework for the SAM system’s operational logic.

AutoSAM utilizes multi-modal retrieval-augmented generation to automate input file creation for the System Analysis Module (SAM) code.

Constructing accurate input files for complex system-level simulations remains a significant bottleneck in nuclear reactor design and safety analysis. This challenge is addressed in ‘AutoSAM: an Agentic Framework for Automating Input File Generation for the SAM Code with Multi-Modal Retrieval-Augmented Generation’, which introduces an agentic framework capable of automatically generating these inputs from diverse, unstructured engineering documentation. By combining large language models with retrieval-augmented generation and specialized document processing tools, AutoSAM extracts relevant parameters and synthesizes validated, solver-compatible input decks for the System Analysis Module (SAM). Could this approach pave the way for prompt-driven reactor modeling, enabling analysts to focus on system definition rather than laborious input file creation?

Streamlining Simulation: Addressing the Input Bottleneck

The creation of input files for sophisticated thermal-hydraulics simulations, such as those performed with the System Analysis Module (SAM), presents a substantial challenge to nuclear engineering workflows. These files, which define the parameters and conditions for modeling complex systems, are not automatically generated and instead demand considerable time and expertise to construct. The process frequently involves manually interpreting detailed engineering documentation and translating it into the specific format required by the simulation code, a task susceptible to human error. Consequently, this manual input creation is not only labor-intensive and costly, but also represents a critical bottleneck that can significantly delay design iterations and thorough safety analyses. The expense extends beyond labor, encompassing the potential for costly errors requiring rework and validation, ultimately impacting project timelines and budgets.

The creation of accurate simulations for nuclear systems is often hampered not by the computational power of the codes themselves, but by the painstaking process of preparing input data. Currently, translating complex engineering documentation – detailing reactor geometry, material properties, and operational parameters – into a format usable by thermal-hydraulics codes like SAM demands significant expertise and manual effort. This reliance on specialized knowledge introduces a critical bottleneck, slowing down design iterations and increasing the potential for human error. Consequently, the time and resources dedicated to input file creation can often exceed those spent on the simulation runs themselves, ultimately limiting the speed and efficiency of nuclear engineering workflows and potentially impacting project timelines and costs.

An agent successfully extracts one-dimensional structural assembly model (SAM) component geometry and connectivity-including labeled nodes, start points, orientations, lengths, and loop flow paths-from a two-dimensional schematic for consistency checking.
An agent successfully extracts one-dimensional structural assembly model (SAM) component geometry and connectivity-including labeled nodes, start points, orientations, lengths, and loop flow paths-from a two-dimensional schematic for consistency checking.

AutoSAM: An LLM-Agent Framework for Automated Input

AutoSAM is an LLM-agent framework engineered to automate the creation of Segmented Attribute Model (SAM) input files. The system accepts diverse engineering documentation formats as input, including text, tables, and diagrams, and processes this heterogeneous data to construct the necessary parameters for SAM simulations. This automation is achieved through an agent-based architecture, eliminating the need for manual translation of engineering specifications into the structured input format required by the SAM modeling environment. The framework is designed to improve efficiency and reduce errors associated with manual input file creation, allowing engineers to focus on model analysis rather than data preparation.

AutoSAM’s operational structure is built upon a ReAct (Reason + Act) agent architecture. This approach enables the system to perform tasks by alternating between reasoning steps, where the Large Language Model (LLM) Agent generates textual rationales, and acting steps, where the agent utilizes external tools. Iterative prompting guides this cycle; the LLM Agent receives a prompt, generates an action (tool use), observes the tool’s output, and then uses this observation to inform subsequent reasoning and actions. This closed-loop process allows AutoSAM to dynamically adapt its behavior and effectively address the task of generating simulation input files from source documentation.

AutoSAM’s performance is enhanced through Retrieval-Augmented Generation (RAG), a technique where the LLM agent accesses and incorporates information from external knowledge repositories during the inference process. This allows the agent to dynamically supplement its internal knowledge with relevant data, improving the accuracy and completeness of generated simulation inputs. Specifically, testing has demonstrated that AutoSAM achieves 100% utilization of all structured parameters provided in the input documentation, indicating a robust capability to translate available data into actionable simulation configurations.

An LLM agent successfully transformed a structured document into a functional simulation model, as demonstrated by the corresponding simulation results.
An LLM agent successfully transformed a structured document into a functional simulation model, as demonstrated by the corresponding simulation results.

Decoding Engineering Data: Multi-Modal Document Processing

AutoSAM utilizes multi-modal document processing to ingest and interpret engineering data from a variety of source formats. This capability extends beyond simple text extraction to include the analysis of Portable Document Format (PDF) files, raster images, and structured spreadsheet data. The system is designed to handle documents regardless of their original creation method or layout, employing specialized tools for each format to ensure data integrity. This approach facilitates the automated capture of information typically found in engineering drawings, reports, and calculation sheets, eliminating the need for manual data entry and reducing the potential for errors.

AutoSAM utilizes a suite of dedicated tools to facilitate data extraction from varied engineering document types. The PDF Analysis Tool employs optical character recognition (OCR) and layout analysis to identify and extract text and tabular data from PDF files, accommodating both native text and scanned images. The Image Analysis Tool focuses on extracting data embedded within images, including dimensions, symbols, and text, through image processing techniques. Finally, the Excel File Reader directly parses data from spreadsheet files, handling numerical values, text strings, and formulas. These tools work in concert to convert unstructured data into a standardized format for downstream processing, supporting a wide range of engineering data sources.

AutoSAM employs an Intermediate Structured Representation (ISR) as a critical step prior to generating the final input file, providing a complete, human-auditable record of the data transformation process. This ISR facilitates verification of data lineage and ensures transparency in the automated extraction process. Performance metrics indicate an 88% recall rate when extracting data from PDF documents, demonstrating the framework’s efficacy in identifying and capturing crucial information required for downstream simulation applications. The ISR’s structured format allows for efficient data validation and debugging, improving the overall reliability of the automated workflow.

Agents leverage a dynamic user prompt and a suite of tools to process multi-modal documents.
Agents leverage a dynamic user prompt and a suite of tools to process multi-modal documents.

Adaptability and Scalability: A Modular Approach to Simulation

The AutoSAM framework employs a deliberately tiered architecture, differentiating between tools that operate independently of the underlying simulation code and those requiring specific adjustments. Solver-Agnostic Tools, representing the foundational layer, provide general functionalities – such as geometric processing or data management – applicable across diverse simulation environments without modification. Conversely, Solver-Specific Tools necessitate adaptation to align with the unique syntax and requirements of each target solver. This modular approach not only enhances the framework’s versatility but also streamlines integration with existing simulation codes, like SAM, by isolating the components requiring solver-dependent adjustments and facilitating a more efficient workflow for nuclear simulation tasks.

AutoSAM’s architecture is intentionally built upon a modular design, a key factor in its ability to scale and integrate seamlessly with diverse simulation codes, notably including the Systems Analysis Module (SAM). This approach separates the framework’s core functionalities into independent, reusable components, allowing for easy adaptation to different computational environments without requiring extensive code modification. The result is a highly flexible system capable of accommodating new solvers and simulation types as they emerge, ensuring long-term viability and broadening its applicability across a range of nuclear engineering applications. This adaptability significantly reduces the barriers to entry for utilizing advanced simulation tools and fosters collaboration between researchers employing varied computational methods.

AutoSAM dramatically streamlines nuclear simulation workflows through automated input file generation, substantially lowering both the time and financial resources required for complex analyses. The framework’s capacity to accurately extract 37 geometric attributes from engineering diagrams with 100% completeness allows for rapid prototyping and iterative design improvements. This capability not only accelerates the simulation process but also facilitates more exhaustive safety assessments, ultimately contributing to enhanced reactor designs and operational reliability. By minimizing manual effort and maximizing data fidelity, AutoSAM empowers engineers to explore a wider range of design options and conduct more rigorous evaluations, leading to safer and more efficient nuclear power systems.

This test case demonstrates an LLM agent successfully converting an unstructured document into structured data, a solver input file, and ultimately, corresponding simulation results.
This test case demonstrates an LLM agent successfully converting an unstructured document into structured data, a solver input file, and ultimately, corresponding simulation results.

The presented framework, AutoSAM, embodies a reductionist principle in its approach to nuclear reactor modeling. It distills complex engineering documentation into manageable input files, prioritizing essential information for simulation. This aligns with Bertrand Russell’s observation that “The point of education is to teach people to think, not to memorize.” AutoSAM doesn’t simply replicate data; it interprets and structures it, mirroring the cognitive process of understanding. The system’s Retrieval-Augmented Generation (RAG) component exemplifies this – extracting only relevant details, avoiding unnecessary complexity, and thereby reducing the cognitive load on both the system and the simulation user. Unnecessary information is, indeed, violence against attention.

What Remains?

The elegance of AutoSAM lies not in what it adds to the process of nuclear reactor simulation, but in what it obviates. The framework demonstrates a capacity to navigate the inherent messiness of engineering documentation, translating disparate formats into actionable model parameters. Yet, a system predicated on retrieval-augmented generation remains, fundamentally, a reflection of its sources. The limitations of the underlying knowledge base, the biases embedded within historical designs, these are not problems solved by automation, but amplified by it.

Future iterations will likely focus on the refinement of multi-modal understanding. However, a more critical path involves addressing the quality of the retrieved information. A perfect system for parsing flawed data will only produce flawed results, albeit with increased efficiency. The challenge, then, is not simply to automate input file generation, but to integrate mechanisms for knowledge validation and error detection – to build a system that questions its own assumptions.

Ultimately, AutoSAM’s success invites a re-evaluation of the simulation process itself. If the most time-consuming aspect of modeling is not the computation, but the preparation of the input, then perhaps the focus should shift towards standardizing documentation and improving the accessibility of validated design data. The true reduction in complexity may not lie in more sophisticated algorithms, but in simpler, more consistent source material.

Original article: https://arxiv.org/pdf/2603.24736.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-28 08:32