Simulation-ready CAD path
NeuroCAD geometry is being structured so SDF authoring, meshing/export, validation, and FEM handoff remain one evidence trail instead of separate demos.
NEUROFEM
Physics AI and finite-element acceleration
NeuroFEM
NeuroFEM is the physics-intelligence track around NeuroCAD: simulation-ready geometry, HPC-backed model training, MOOSE validation, and a CutFEM research path for differentiable FEM on SDF geometry.
5,000 CPU/GPU compute hours secured
Initial allocation for foundation-physics model training and validation.
8 / 9 release gates passed
The HPC evidence pack reports eight gates passed and one intentionally marked not_ready.
4 physics regimes unified
Thermal, species, thermo-mechanical, and thermo-mechanical-species fields in one model path.
0.35% K_I CoV evidence
Crack submodel evidence from live MOOSE validation in the execution ledger.
WHAT IS BUILT
Compute campaign package
The NeuroFEM HPC pack documents EuroHPC-track submissions, a defensible 5,000-hour minimum, and a 15,000-hour expansion ask.
Unified multiphysics training loop
Current 32-cubed evidence shows one unified model learning four regimes without catastrophic negative transfer; longer runs exposed data limits rather than simple training-time limits.
MOOSE crack submodel evidence
Task-2 validation has live-MOOSE evidence: boundary condition check, stable K_I sampling, dispatcher coverage, and cluster smoke output with finite K_I.
Differentiable CutFEM research trail
The branch evidence trail contains CutFEM tests, a differentiable CutFEM-on-SDF thesis draft, and mesher planning documents for a four-vector differentiable mesher.
Evidence pack instead of slides
The internal pack keeps CLI health, release gates, scenarios, git head/log, and codebase metrics together so claims can be checked before they are used externally.
EVIDENCE TRAIL
HPC/INDEX.md
Applicant facts, compute ask, honest framing, EuroHPC-track package, and reproducible evidence pack index.
HPC/52_EXECUTION_LEDGER.md
Current training truth: unified_v1 is best, unified_v2 overfit, resume-RNG bug fixed, MOOSE crack workflow proven.
HPC/56-59 in git history
Differentiable mesher, connected training plan, frontier research plan, and development plan.
tests/test_cutfem/
Branch trail includes eight CutFEM-focused test files covering heat, hybrid, solver, sparse unified, 3D, thermoelastic, and backend behavior.
docs/differentiable-cutfem-on-sdf-thesis.md
Publishable thesis-style manuscript from abstract through references.
CURRENT TRUTH
The page is intentionally evidence-first.
NeuroFEM is strong enough to discuss with strategic engineering partners, but it is framed as a validation program. The next milestone is customer evidence, not a broad production launch.
- No public claim of finished production accuracy.
- No claim that a foundation physics model is already trained.
- Current results show that more data, regularization, and early stopping matter more than just longer training.
- The near-term external value is pilot/customer validation evidence: geometry workflow, simulation handoff, and measurable technical feedback.
Pilot evidence target