Beyond Simulation

Engineering simulation today follows a familiar pattern: you design a part, mesh it, apply loads and boundary conditions, run a solver, inspect the results, and manually adjust the geometry. Repeat until the design converges — or until time runs out.

This is simulation as analysis. It tells you whether a given design works. It does not tell you what a better design would look like.

NeuroFEM is simulation as synthesis. It does not merely evaluate your geometry — it invents geometry that satisfies your performance requirements. Given a design space, a set of load cases, and target constraints (maximum stress, minimum stiffness, volume fraction), NeuroFEM directly optimizes the underlying signed distance field to produce an optimal shape — in milliseconds, not hours.

This is not incremental improvement over traditional FEA. This is a different category of engineering tool.

The Architecture: Foundation Model + Physics Router

NeuroFEM is built around a mechanical engineering foundation model — a large neural network trained from scratch on mechanical physics. Unlike general-purpose AI models that learn from internet text or images, this model learns from physics: stress distributions, displacement fields, vibration modes, thermal gradients, and contact pressures.

The architecture is based on Transolver Physics-Attention, a modern transformer variant designed specifically for physical field prediction. Scaled to 1–10 billion parameters depending on compute budget and performance targets, the foundation model internalizes mechanical physics across structural, modal, thermal, contact, and fatigue domains.

Critically, NeuroFEM operates through NeuroCAD’s PhysicsExpert router, not as a replacement for all simulation. The router maintains a registry of solver capabilities:

Authority Class	Role	Example
Approximate	Fast preview, design exploration, optimization	Foundation model, classical surrogate, analytical formula
Authoritative	Final validation, certification, regulatory submission	MOOSE FEA, validated third-party solver

The foundation model provides Approximate predictions — fast enough for design exploration, accurate enough to guide decisions, but never treated as certification evidence. When a design approaches finalization, the router escalates to MOOSE or an equivalent authoritative solver for validation.

This separation is architectural law in NeuroCAD. AI proposes. Physics validates.

Inverse Topological Design

The most radical capability of NeuroFEM is inverse design — working backwards from performance requirements to geometry.

Traditional topology optimization iterates: guess a density distribution, run FEA, compute sensitivities, update densities, repeat. Each iteration requires a full finite element solve. On a complex 3D part, this can take hours per iteration, and hundreds of iterations to converge.

NeuroFEM’s foundation model inverts this loop. Because it has internalized the relationship between geometry and physical response across millions of training examples, it can directly predict the geometry that satisfies a given set of performance requirements. This is implemented through Deep Equilibrium Models (DEQ) running on JAX/XLA:

The design space, loads, and constraints are encoded as inputs.
The DEQ finds an equilibrium state — a geometry whose physical response matches the requirements.
The equilibrium state is decoded into an SDF — the native geometric representation of NeuroCAD.
The field gradients flow back through the DEQ, enabling gradient-based refinement.

The result: a topology-optimized part in milliseconds, with gradients available for further refinement. What takes hours in traditional TopOpt takes seconds in NeuroFEM.

Because the output is an SDF — not a mesh, not a density field — it is immediately usable in NeuroCAD for further editing, lattice infill, material grading, or manufacturing validation. There is no reconstruction step, no surface fitting, no loss of fidelity.

The Autonomous Data Factory

A foundation model is only as good as its training data. NeuroFEM solves the data problem by building its own.

Using Active Learning (Simula) combined with NeuroCAD’s Rust-based Distributed Constraint Graph engine, NeuroFEM operates an autonomous data generation pipeline:

Generate. The DCG engine produces millions of diverse geometric configurations — different part shapes, load cases, boundary conditions, and material properties — parametrically and deterministically.
Solve. Each configuration is sent to MOOSE for authoritative finite element solution. The solver produces stress tensors, displacement fields, strain energy distributions, and modal frequencies.
Learn. The foundation model trains on the growing dataset. Active learning identifies regions of the design space where the model’s uncertainty is highest — and generates more training samples in those regions.
Repeat. The model improves with every cycle. It never makes the same mistake twice because every error becomes a targeted training example in the next iteration.

The target scale: approximately 170,000 per-family DOE simulations across six mechanical element families, plus 500,000 freeform geometry simulations from open-source geometry corpora, plus 10,000 benchmark and reference problem solutions. All generated, solved, and ingested autonomously.

This is not a one-time training run. This is a continuous learning system that improves as it operates.

Confidence Gating: Knowing What You Don’t Know

AI predictions are only useful if you know when to trust them. NeuroFEM embeds confidence quantification at every level of its architecture:

Confidence Range	Action
c ≥ 0.85	Accept approximate prediction as primary exploration result
0.70 ≤ c < 0.85	Accept with marginal badge; MOOSE validation queued
0.40 ≤ c < 0.70	Reject or try alternative expert; escalate to MOOSE
c < 0.40	Mandatory MOOSE escalation
Out of Distribution	Confidence clamped low; mandatory escalation

This gating is not optional. It is enforced by the router at the architectural level. A prediction with confidence below 0.85 is never presented as reliable without qualification. A prediction below 0.40 is never presented at all — the system escalates to authoritative solving instead.

For regulated industries — aerospace, medical devices, automotive safety — this gating provides a defensible boundary between AI-assisted exploration and engineering certification. The AI helps you explore the design space faster. The authoritative solver provides the evidence for certification.

The Three-Level Router

The PhysicsExpert router operates at three levels of intelligence:

Level 1 — Dispatcher. Ranks candidate experts (foundation model, classical surrogate, analytical formula, MOOSE) by capability match, estimated latency, historical accuracy, current load, and authority requirements. Selects the best expert for each task.

Level 2 — Steerer. Adapts expert weights over time based on calibration drift, prediction latency, user retry/acceptance patterns, and tenant-specific history. A heat exchanger designer and a bracket designer may route through different experts because their physics domains differ.

Level 3 — Corrector. When an approximate prediction is close but not converged, the corrector refines it through residual PINN (Physics-Informed Neural Network), Newton iteration, conjugate gradient, or multigrid smoothing. This turns a “good enough” prediction into a converged solution without the cost of a full authoritative solve.

Multiphysics: Beyond Structural Analysis

NeuroFEM is not limited to structural mechanics. The foundation model handles coupled physics simultaneously:

Structural: Stress, strain, displacement, contact pressure, fatigue life
Thermal: Temperature distribution, heat flux, thermal stress, convection boundaries
Modal: Natural frequencies, mode shapes, harmonic response, damping ratios
Fluid-Structure: Coupled thermal-structural, pressure-structural interaction

Internal Expert-Choice routing networks dynamically specialize neural reasoning based on the physics domain of each task. A structural-only problem routes through structural-specialized subnetworks. A coupled thermal-structural problem activates both subnetwork sets and their interaction pathways. The model learns which physics interact and how — without explicit coupling code.

Generative Design: Optimization Meets SDF

The integration between NeuroFEM and NeuroCAD’s SDF substrate creates capabilities that neither could deliver alone:

Level-set topology optimization on SDF — the optimizer directly modifies the signed distance field, with gradients provided by NeuroFEM’s foundation model. No mesh, no remeshing, no density interpretation.
Differentiable lattice authoring — lattice parameters (cell type, size, wall thickness, density gradient) are exposed as differentiable variables. NeuroFEM optimizes them against structural or thermal objectives, producing functionally graded lattices tuned for specific performance targets.
Multi-physics objective handling — a single optimization run can balance structural compliance, thermal conductivity, vibration avoidance, and volume fraction. The router composes multiple experts into a unified objective function.
Constraint-aware optimization — manufacturing constraints (minimum wall thickness, overhang angles, draft requirements) are compiled into the optimization problem. The optimizer cannot propose geometry that violates active manufacturing rules.

Every generative iteration produces a verifiable receipt — a NeuroGraph patch recording inputs, expert selections, confidence scores, and results. This audit trail is essential for regulated industries and provides traceability from requirements to final geometry.

From Research to Production

NeuroFEM is shipping in lockstep with NeuroCAD’s phase roadmap:

Phase 1–2 (Core + Features): Simulation workbench for task authoring. Manual mesh preparation. MOOSE integration for authoritative structural and thermal FEA. Result visualization and comparison.

Phase 3 (Organic + Optimization): Foundation model deployment as Approximate expert. Inverse design for structural topology optimization. Design space explorer for parameter sweeps and trade studies. Confidence gating active.

Phase 4 (Production): Full autonomous data factory operational. Multi-physics coupled optimization. Generative design with manufacturing constraints. Benchmark suite for regulatory qualification. Publication and IP strategy execution.

The Engineering Implication

NeuroFEM changes what it means to simulate:

From hours to milliseconds. The foundation model predicts physical fields in a single forward pass. What takes a traditional solver hours of CPU time takes NeuroFEM milliseconds on a GPU.
From analysis to synthesis. You no longer ask “does this design work?” You ask “what design satisfies these requirements?” and the system produces geometry — not just pass/fail verdicts.
From manual to autonomous. The data factory generates its own training data. The model improves continuously. The confidence gates ensure safety during this evolution.
From disconnected to integrated. NeuroFEM operates directly on NeuroCAD’s SDF geometry. There is no export-import-mesh-solve-remodel loop. The simulation lives inside the design environment.

NeuroFEM does not replace FEA. It augments it — handling the exploration, the iteration, the optimization, and the parameter sweeps that consume engineering time, while reserving authoritative solving for final validation. The engineer’s role shifts from running simulations to steering optimization. The AI does the computation. The engineer makes the decisions.

NeuroFEM: The AI That Invents Optimal Geometry