Proof of Concept Plan

The goal is to develop a compelling proof of concept to secure pre-seed funding, followed by a seed-stage machine learning system built on the infrastructure it establishes. The plan splits into two phases: a pre-seed demo based on heuristic model-aware meshing (no ML required), and a seed milestone that replaces the heuristic with a trained graph neural network policy using data generated by the pre-seed phase.

See FEA Sidekick Plan for the product vision and Founder Background and Market Context for the market rationale.

Prerequisites

Before any development begins, resolve the following:

• GPL 2 licensing strategy: CalculiX is GPL 2 licensed. Any modification to its source must be open-sourced. The preferred approach for the PoC is to interface with CalculiX at the input/output level only — invoking it as an unmodified subprocess — which avoids triggering the GPL 2 copyleft requirement. Tighter solver integration (e.g., Jacobian-level hooks) is a post-partnership milestone and requires a separate legal position. Confirm this before writing any code that touches CalculiX internals.
• University IP clearance: Development must not begin until obtaining a written NoC from Cranfield University confirming that no employer resources, data, or IP are involved. See Founder Background and Market Context for the broader legal context.

Open Source Contributions

The startup will contribute all tree-sitter grammars for FEA solvers developed during this project to open source in the long term. Contributions may be held back during stealth mode; the legal pathway for this requires further exploration.

Tool Stack

• pygccx — Python API for building CalculiX models programmatically, enabling automated model generation and systematic mesh variant creation for the data pipeline.
• gmsh Python API — mesh generation and mesh refinement, driven by size fields produced by the model graph.
• py-tree-sitter — parsing CalculiX input decks into syntax trees for semantic extraction.
• PyTorch + PyTorch Geometric — GNN model training and graph-structured mesh learning. PyTorch Geometric is the standard library for mesh and graph neural networks and is the chosen framework for the seed-phase ML work.
• CalculiX — FEA solver, invoked as an unmodified subprocess.
• Cloud compute — automated simulation runs for the data pipeline and model training (seed phase).

Main Pitch

The proof of concept must satisfy the following goals to demonstrate a sound product:

• Accelerated simulations that preserve mathematical rigour.
• AI positioned as the next step in scientific computing, rather than as an all-knowing oracle.

Demo Problem

The demo geometry is a simplified WAAM deposited wall with linear elastic boundary conditions (single load step, no thermal coupling). This choice is deliberate on two grounds:

• Geometric relevance: WAAM components are a known target application for both GE Aerospace and Airbus. A demo on this geometry will be immediately recognisable and credible to simulation engineers at those organisations in a way that a textbook cantilever beam is not.
• Tractable physics: Restricting to linear elastic with a single load step keeps simulation times short and results interpretable for the PoC, while the geometry retains the mesh challenge that makes the problem interesting: stress concentrations at layer interfaces and substrate transitions are poorly captured by default isotropic meshing, which gives the model-aware size field a clear, demonstrable target.

Geometry from existing Cranfield WAAM work can be used directly, with BCs simplified for the PoC. This avoids building a new geometry from scratch.

Benchmark Definition

The baseline is a default Gmsh mesh of the demo geometry with no size field customisation. Improvement is measured against:

• DOF count — reduction while maintaining solution accuracy within a defined tolerance (e.g., <1% error in peak von Mises stress vs. a fine reference mesh).
• Time to convergence — wall-clock time for the CalculiX solve.
• Iteration count — for non-linear extensions in the seed phase.

The reference solution for accuracy comparison is a uniformly fine mesh run. All three metrics must improve or hold for the demo to be considered successful.

Data Pipeline

The automated simulation infrastructure built in Phase 1 is the data pipeline for Phase 2. Each automated run produces a (mesh configuration, solution quality) pair. By sweeping size field parameters and geometry variants, a labelled dataset is generated without additional effort. This means the Phase 1 infrastructure is not throwaway work — it is the foundation for the seed-phase GNN training.

Phase 1: Pre-Seed Demo (Weeks 1–8)

Goal: Demonstrate that parsing the model deck and using its semantic content to drive heuristic mesh refinement produces a measurable improvement on a geometry relevant to the target customer. No ML required.

1. Automate CalculiX model creation for the WAAM demo geometry via pygccx. Establish the benchmark baseline (default Gmsh mesh, baseline metrics).
2. Develop a tree-sitter parser for CalculiX input files.
3. Construct the model graph: link parsed model semantics (sets, surfaces, BCs, loads, materials, contacts) to mesh connectivity via the gmsh Python API.
4. Implement heuristic size field generation from model graph features — refine at load introduction regions, constraint boundaries, and layer interface zones; coarsen in low-gradient interior regions.
5. Close the loop: remesh with the heuristic size field, rerun with CalculiX, and quantify improvement against the benchmark baseline.

Phase 2: Seed Milestone (Post-Funding)

Goal: Replace the heuristic size field with a trained GNN refinement policy. Validate that the learned policy matches or exceeds the heuristic on held-out geometries, including harder problems (contact, non-linear, thermo-mechanical).

1. Use Phase 1 automated runs to generate labelled training data by sweeping size field parameters and geometry variants.
2. Train a GNN-based refinement policy using PyTorch Geometric on the (model graph, mesh configuration, solution quality) dataset.
3. Replace the heuristic size field with GNN recommendations; benchmark against both the baseline and the heuristic result.
4. Extend validation to non-linear and thermo-mechanical models as the policy generalises.