FEA Sidekick
This document outlines a proof of concept for an AI “sidekick” designed for finite element analysis (FEA). The primary focus is accelerating simulation workflows through model-aware meshing, which intelligently allocates computational resources where they matter most.
The long-term product goal is to work with most commercial solvers by converting their model decks into an open, solver-agnostic representation. This intermediate format will enable seamless integration across different simulation environments without vendor lock-in.
See Founder Background and Market Context for the market rationale and the domain expertise underpinning this plan.
Product Scope
- Most AI-based solvers aim for end-to-end prediction, which highly regulated industries will not accept as final validation. Engineers in aerospace and automotive sectors require traceable, physics-based solutions that meet strict certification standards.
- The near-term wedge is model-aware adaptive meshing: using the model deck to understand where accuracy matters, then refining or coarsening to reduce wasted degrees of freedom (DOFs) without sacrificing solution quality.
- By using tree-sitter to parse input files, the long-term product becomes an end-to-end sidekick that accelerates model development and solution time by catching errors early and guiding meshing decisions.
- The long-term goal is to partner with solver developers to embed deeper acceleration techniques (e.g., Jacobian/tangent initialization or other iteration-level methods) directly into the solver for gains unattainable through adaptive remeshing alone.
Proof of Concept Plan
The proof of concept establishes the technical foundation through four key milestones:
- Parse CalculiX/Abaqus-style input decks (via tree-sitter) into an open model format that captures both syntax and semantics.
- Build a linked “model graph” that connects syntax and semantics (sets, surfaces, steps, BCs/loads, materials, contacts) to the underlying mesh connectivity.
- Produce adaptive meshing recommendations (refine/coarsen targets, size fields) that prioritize the regions that matter for the model: load introduction points, constraint boundaries, contact interfaces, and material transitions.
- Close the loop with CalculiX: remesh, rerun, and quantify gains on a small benchmark suite (time-to-QoI, iteration count, DOFs) versus baseline meshing.
Future Path
Once the proof of concept demonstrates viability, the product roadmap includes:
- GNN-based refinement policies and learned error indicators for adaptive remeshing, trained on the data generated during the PoC phase.
- Converters for additional commercial solver formats (Nastran, Ansys) built on the same open model intermediate representation (IR).
- A data factory for high-volume, high-quality training data: automated model variants, multi-fidelity meshes, and systematic solver runs to support robust generalization across problem classes.
- Strategic partnerships with commercial solver developers to integrate iteration-level acceleration where supported, moving beyond pre-processing into solver internals.