A Constrained Natural-Language Interface for Variational Multi-Physics Finite Element Simulations in FEniCS

Researchers present a constrained natural-language interface for finite element simulations that uses LLMs only for front-end parsing tasks while delegating critical solver logic to human-written templates. The system achieves 100% parse validity and demonstrates effective integration of language models with scientific computing by limiting AI to non-critical paths, reducing reliability risks.

This research addresses a fundamental challenge in scientific computing: integrating large language models without compromising numerical accuracy or solver reliability. Rather than allowing LLMs to generate solver code directly—a risky approach when computational correctness is critical—the authors implement a carefully constrained architecture where natural language processing handles only input parsing and geometry generation, while a deterministic dispatcher maps validated specifications to pre-validated FEniCS templates.

The approach reflects broader industry maturation in AI-assisted software development. As organizations adopt LLMs for code generation, the critical insight here is that autonomy and reliability often conflict. By restricting LLM involvement to front-end tasks with built-in retry mechanisms, the system maintains numerical integrity while still reducing manual setup effort. The validation results support this philosophy: the parser achieves 100% final parse validity through constrained design, while the numerical templates match analytical benchmarks within 0.5-5% depending on problem complexity.

For scientific computing and engineering software, this architecture offers a practical blueprint for AI integration without sacrificing trust. The system successfully demonstrated a complex real-world scenario—3D elastoplastic analysis with geometric features—from a single natural-language prompt, suggesting genuine utility for domain experts. This differs markedly from open-ended code generation approaches that prioritize capability over verifiability.

The implications extend beyond finite element analysis. Industries requiring high-consequence numerical computation—aerospace, structural engineering, materials science—could adopt similar patterns to leverage LLM productivity gains while maintaining regulatory and safety compliance. Future work will likely explore expanding template libraries and automating template validation, but the core principle of constraining AI to non-critical paths appears robust.

• →LLMs handled only parsing and geometry generation, with all critical solver logic delegated to validated human-written templates.
• →The system achieved 100% final parse validity and 97.1% field-extraction accuracy through constrained design and retry mechanisms.
• →Numerical validation showed sub-percent agreement on smooth cases and 2-5% accuracy on nonlinear benchmarks compared to analytical solutions.
• →Successfully demonstrated practical utility by generating and analyzing a 3D elastoplastic bracket with geometric features from natural language.
• →Presents a generalizable architecture for integrating LLMs into high-stakes computational software without sacrificing reliability.