PaperlayerView exact revision on arXiv

Source-linked AI summary

A Multi-AI-agent Framework Enabling End-to-end Finite Element Analysis for Solid Mechanics Problems

Titu Ranjan Sarker, Muhammed Jawaad Zulqernine, Ling Yue, Shaowu Pan, Chenxi Wang, Shiyao Lin

arXiv:2606.00138v1cs.AI

TL;DR

Real-world finite element analysis requires substantial interdisciplinary knowledge and engineering experience, while careless use can produce serious consequences. AbaqusAgent uses six AI agents to automate end-to-end Abaqus simulations from natural-language prompts and achieved an 86% overall success and accuracy rate across 50 validation cases.

  • Problem

    Real-world finite element analysis requires substantial interdisciplinary knowledge and engineering experience, and careless black-box use may lead to catastrophic consequences.

  • Method

    AbaqusAgent uses six orchestrated agents to interpret requirements, retrieve analogous cases, construct and execute simulations, correct errors, and visualize results.

  • Results

    86% overall success and accuracy rate was achieved across 50 validation cases.

  • Takeaways & Limitations

    AbaqusAgent showcases the potential to increase accessibility and lower entry barriers for engineering simulation and computational mechanics education.

  • Takeaways & Limitations

    The framework's procedure assumes a maximum number of iterations for producing the final simulation configuration.

Abstract

from arXiv · show

Finite element analysis (FEA) is the most important numerical approach for solid mechanics. Challenges of FEA include a steep learning curve for entry-level users and potential false simulations due to incorrect definitions of key simulation components, such as boundary conditions, load cases, and solution variables. Years of engineering experience are usually necessary for real-world problem-solving. To address these issues, we present AbaqusAgent, a multi-agent framework grounded in large language models (LLMs) for solid mechanics analyses. AbaqusAgent is developed to facilitate analysis case generation and execution using Abaqus, one of the most widely used FEA packages, by turning users' natural-language instructions into executed FEA analyses and result visualization. AbaqusAgent is composed of six agents, including interpreter, architect, input writer, runner, reviewer, and visualizer agents, encompassing all the essential pre-processing and post-processing steps of standard FEA analyses. A wide variety of 50 solid mechanics problems have been successfully validated, achieving an overall success rate of 86%. Beyond improving the efficiency of FEA for solid mechanics problems and lowering the barrier to computational mechanics education, AbaqusAgent advances the human-simulation interaction paradigm and enables integration with AI-empowered optimization and material characterization workflows. The code is available at https://github.com/LIRAM-LIN/AbaqusAgent

1. Introduction

The introduction positions FEA as powerful but expertise-intensive, with careless use potentially causing catastrophic consequences, while existing API-based automation remains constrained by predefined scripts. It presents AbaqusAgent as a six-agent framework designed to automate Abaqus simulations, supported by a 104-problem RAG repository and evaluated across 50 solid mechanics problems.

  • Motivation: FEA supports modern engineering design validation and digital twins, but real-world application requires interdisciplinary knowledge and substantial engineering experience.Careless or “black-box” FEA usage may lead to catastrophic consequences.
  • Related work: API-based FEA automation enables structural studies, topology optimization, geometry generation, and damage investigations but often lacks flexibility because it depends on predefined scripts.The passage identifies Abaqus-Python API applications across several engineering tasks.
  • Focus of this work: AbaqusAgent aims to run FEA simulations automatically for engineering problem-solving and education through agent orchestration, a versatile Abaqus RAG database, and evaluation of accuracy, efficiency, and token usage.These are the three major aspects identified for the framework.
  • Main contributions: 86% success rate was achieved across 50 solid mechanics problems using AbaqusAgent’s six-agent architecture.The architecture is described as orchestrated agents for solving a wide variety of solid mechanics problems.
  • Main contributions: 104 solid mechanics problems were curated in a heterogeneous RAG repository paired with hierarchical scoring-based search for similar cases and Abaqus input-file templates.The search strategy helps the LLM follow and modify retrieved cases and templates.

2. Methodology

AbaqusAgent implements an end-to-end, six-agent workflow that converts natural-language requirements into Abaqus simulations and result visualizations. Its methodology combines requirement structuring, hybrid retrieval from 104 curated cases, input generation, execution, iterative error correction, and post-processing.

  • End-to-end architecture: The six-agent workflow interprets requirements, retrieves analogous cases, writes and runs Abaqus inputs, reviews failures, and visualizes successful results.The agents are Interpreter, Architect, Input Writer, Runner, Reviewer, and Visualization.
  • Requirement interpretation: The Interpreter Agent extracts geometry, materials, boundary conditions, loading, and requested outputs, requesting missing parameters before modeling.It also paraphrases the prompt for the Architect Agent.
  • Case retrieval: The hybrid search weights case name, category, and material at 30%, 60%, and 10%, respectively, after filtering candidates by analysis domain.FAISS first retrieves candidate cases, which are then filtered and re-ranked using the structured metadata.
  • Iterative refinement and visualization: Failed simulations trigger Reviewer Agent corrections sent back to the Input Writer Agent, with up to 15 iterations before a case is considered unsolved.Successful analyses are forwarded for Python-based visualization of deformation values and contour plots.
  • Case retrieval: 104 curated solid mechanics problems support retrieval, with candidates selected by semantic similarity, exact domain filtering, and weighted ranking.The database contains 71 Abaqus benchmark-manual problems and 33 textbook-style problems.

3. Results

Across 50 benchmark cases, AbaqusAgent achieved 86% simulation success and 86% result accuracy, with RAG improving both performance and efficiency. The framework also handled varied geometries and nonlinear simulations, while reviewer use, underlying LLM capability, and prompt specificity materially affected outcomes.

  • Overall performance: 86% simulation success and 86% result accuracy were achieved across 50 evaluation cases.The dataset comprised 40 modified RAG cases and 10 external cases.
  • Overall performance: 92.5% success with RAG coverage exceeded 60% outside coverage, while retrieval accuracy reached 85% for inside-RAG cases.Among 40 inside-RAG cases, 34 identified the most similar case.
  • RAG efficiency: 28,761 tokens and 157 seconds were used inside RAG coverage, compared with 48,200 tokens and 312 seconds outside coverage.These averages were reported for successful cases and indicate lower token consumption and solution time when relevant examples and input files were available.
  • Geometry versatility: The framework modeled circular, elliptical, and square-hole tension geometries, including geometries absent from the RAG benchmark.The circular-hole case was included in the benchmark, whereas the elliptical and square-hole geometries demonstrated modeling of previously undocumented geometries.
  • Modeling capabilities and prompt effects: Prompt specificity affected open-hole plate accuracy, while the framework also handled nonlinear dynamic impact and elastic–plastic deformation cases.Omitting meshing instructions produced an inaccurate unstructured mesh, whereas a detailed prompt specifying a radial mesh yielded a well-defined mesh.
  • Ablation study: 90% success and 90% result accuracy were obtained with both RAG and the Reviewer Agent enabled, whereas disabling the reviewer reduced both rates to 45%.With RAG and reviewer enabled, retrieval accuracy was 80%; without the reviewer, retrieval accuracy remained 80%.

4. Conclusions

AbaqusAgent automates end-to-end physics-based FEA from natural-language prompts across diverse solid mechanics problems. The framework broadens access to engineering simulation, while currently lacking multiphysics and custom-mesh capabilities.

  • Contributions: AbaqusAgent automates end-to-end physics-based FEA simulations using natural-language prompts.The framework is designed for a wide variety of solid mechanics problems.
  • Scope: The framework covers solid mechanics problems involving stressing, deforming, yielding, cracking, impacting, vibrating, and buckling.
  • Limitations: AbaqusAgent currently does not support multiphysics or custom-mesh solid mechanics simulations, which are deferred to future work.
  • Impact: AbaqusAgent demonstrates the potential of AI-agent frameworks to increase accessibility and lower entry barriers for engineering simulation and other scientific disciplines.

Appendix A. Validation and Extended Capabilities of AbaqusAgent · Appendix A.1. Validation Against Analytical Benchmark

AbaqusAgent was validated against the closed-form buckling solution for a simply supported rectangular plate under uniaxial compression. The finite-element prediction achieved 98.3% accuracy and reproduced the required model definitions without user intervention.

  • Appendix A.1. Validation Against Analytical Benchmark: The framework’s physical correctness was tested against classical buckling of a simply supported rectangular plate with a closed-form analytical solution.The benchmark concerns a plate under uniaxial compression.
  • Appendix A.1. Validation Against Analytical Benchmark: The benchmark used a 1 m × 3 m plate with 2 mm thickness, E = 70,000 N/mm2, and ν = 0.33.All edges were simply supported, and compression acted along the 3000 mm length.
  • Appendix A.1. Validation Against Analytical Benchmark: The analytical reference followed Timoshenko’s classical thin-plate buckling theory, equivalent to Kirchhoff plate theory.The critical buckling load was formulated per unit width.
  • Appendix A.1. Validation Against Analytical Benchmark: For the plate’s aspect ratio a/b = 3, the fundamental simply supported mode used buckling coefficient kc = 4.0.The analytical expression depends on the plate’s flexural rigidity D.
  • Appendix A.1. Validation Against Analytical Benchmark: The AbaqusAgent model used a 10×30 structured S4R mesh with five through-thickness integration points and explicitly imposed support constraints.U3 was restrained along all edges, while corner constraints suppressed in-plane rigid-body motion.
  • Appendix A.1. Validation Against Analytical Benchmark: 98.3% solution accuracy was obtained when AbaqusAgent’s finite-element buckling result was compared with the Kirchhoff-theory solution.The comparison addressed compressive out-of-plane buckling of a slender plate.
  • Appendix A.1. Validation Against Analytical Benchmark: The finite-element result agreed with the analytical solution to within the reported validation discrepancy.The supplied passage does not include the numerical discrepancy value.
  • Appendix A.1. Validation Against Analytical Benchmark: The validation showed that AbaqusAgent generated the geometry, boundary conditions, load definition, and analysis type without user intervention.The remaining discrepancy was attributed to finite-element discretization error during eigenvalue extraction.

Appendix A.2. Motivation for Using Abaqus over alternative FEA Solvers

Abaqus was selected for its mature environment, robust nonlinear capabilities, and standardized syntax, enabling AbaqusAgent to automate natural-language-driven FEA workflows for beginner and expert users.

  • Abaqus provides a mature simulation environment, robust nonlinear analysis capabilities, and well-established input file syntax for benchmark-based validation.
  • AbaqusAgent automates the workflow from natural-language problem statements through simulation execution, error review, and result visualization.
  • Abaqus offers a standardized, widely adopted structural-analysis environment more suitable for automated natural-language workflow generation than expertise-intensive, customizable open-source frameworks such as MOOSE.
  • Commercial solvers such as Abaqus and Ansys are commonly preferred in industry because of their robustness for strongly nonlinear problems including damage, high-speed deformation, and multi-body contact.

Appendix A.3. Details about Abaqus simulation files

Appendix A.3 describes the files produced by Abaqus and the custom workflow to define execution conditions, record solver behavior and numerical results, and support review, debugging, and visualization.

  • Abaqus-generated files: Abaqus automatically generates control, monitoring, diagnostic, and result files that document execution, solver behavior, errors, warnings, and numerical results.These files enable the reviewer agent to diagnose failures, guide corrective actions, and facilitate post-processing and visualization.
  • Workflow-generated files: The custom workflow generates similar_case.txt for retrieved-case information, while AbaqusInput.env defines memory and CPU usage settings.These files support reference, automation, debugging, and management of execution conditions.
  • Core simulation files: AbaqusInput.inp defines the model problem, while AbaqusInput.dat stores detailed numerical output for result verification.The input file may be generated or written by a user or script, whereas the data file stores requested printed data.
  • Core simulation files: AbaqusInput.msg logs solver messages and warnings, and AbaqusInput.rpt stores extracted results in a readable format.These files support understanding solver behavior and accessing extracted analysis results.

Appendix B. Prompt Design for AbaqusAgent

Appendix B specifies expert-designed prompts that guide AbaqusAgent through finite-element workflows, with agent-specific rules for interpreting requirements, naming cases, and generating runnable Abaqus input files. The prompts prioritize preserving user intent, separating physics from syntax templates, and enforcing structured output and validation rules.

  • Appendix B. Prompt Design for AbaqusAgent: Experts develop prompts from Abaqus syntax guidelines and finite-element modeling flow to improve accuracy across the AbaqusAgent workflow.The prompts are intended to guide agents through simulation tasks while improving their handling accuracy.
  • Appendix B.1. Interpreter Agent prompts: The interpreter rewrites raw requests into professional Abaqus-style prompts and checks geometry, material properties, boundary conditions, loading conditions, and requested output.It must preserve technical intent, improve clarity, avoid inventing unspecified details, and return a structured JSON object listing missing items.
  • Appendix B.2. Architect Agent prompts: The architect converts requirements into structured case descriptions whose names reflect analysis type, geometry, and important loading or conditions.The prompt also requires case name, domain, category, and material, while excluding numerical values from case names.
  • Appendix B.3. Input Writer Agent prompts: The input writer generates complete, correct, and runnable .inp files using the user requirement and a retrieved similar-case input file as a structural template.The reference file may guide syntax, ordering, block structure, naming, and reusable implementation patterns, but not the underlying physics.
  • Appendix B.3. Input Writer Agent prompts: The user requirement always overrides the reference case for load type, load direction, boundary conditions, deformation mode, element choice, and analysis procedure.If a reference describes compression but the user requests bending, the generated model must replace compression behavior with bending-consistent modeling.
  • Appendix B.3. Input Writer Agent prompts: Generated mesh syntax such as *NGEN, *NFILL, and *ELGEN should be preserved for regular structured geometries whenever feasible.The prompt discourages manual node- and element-by-element listings when generated definitions remain applicable and makes generated mesh syntax mandatory whenever feasible for similarly structured domains.
  • Appendix B.3. Input Writer Agent prompts: The input-file prompt enforces Abaqus-only output, a mandatory keyword sequence, closed blocks, correct placement of sections, loads, boundary conditions, and outputs, and a minimum mesh of 20 elements.It also prohibits explanations, markdown, extra text, blank lines, pseudo-syntax, and history output unless explicitly requested.

Appendix C. User prompts structure in case studies

Benchmark cases use natural-language prompts to define essential FEA parameters and requested outputs. Supplementary information reduces ambiguity and supports more accurate agent reasoning for complex problems.

  • Each benchmark prompt specifies the analysis type, geometry, material properties, boundary conditions, loading conditions, and requested outputs.Prompts may include supplementary information to minimize ambiguity and improve agent reasoning, especially for complex problems.
Loading 2606.00138v1…