The Intelligence Layer
There are no production-grade systems from Anthropic, OpenAI, or Google DeepMind that readily apply to computational fluid dynamics, finite element analysis, tolerance stack-ups, or failure-mode reasoning.
Zero.
They've collectively produced AlphaFold for protein structure, ChatGPT for everyday chat, Claude Code for coding, and GraphCast for weather prediction. Billions into code generation, biology, and chat. But the hardware engineering disciplines that produce cars, aircraft, satellites, and medical devices, the disciplines where a missed measurement kills people, have no other AI research program behind them.
NIST AMS 100-26 quantifies $8.4 billion per year in spending by U.S. manufacturers on engineering questions that a structured knowledge system could eliminate. Last year, thirteen automakers paid $57 billion in warranty costs alone, and the medical device industry's direct cost of quality was $26–36 billion.
The AI landscape for engineering organizes along two axes: reasoning fidelity (statistical to causal/symbolic) and domain grounding (general corpus to physics-constrained).
The prevailing narrative says this gap is temporary: scale the model, scale the reasoning. Benchmarks like GPQA-Diamond seem to support it. Gemini 3.1 Pro scores 94.3%, a 55-point gain over GPT-4's 2023 baseline.
But when tasks require constructing a model that demonstrates real-world effectiveness rather than merely recalling, the trend breaks down.
On ChipBench, Claude Opus 4.5 achieves 30.74% on Verilog generation and 13.33% on reference-model generation in real chip-design contexts. On LLM-SRBench, across 239 decontaminated equation-discovery problems, LLMs demonstrate memorization rather than effective reasoning. Even in simpler domains, adding irrelevant clauses to grade-school math drops accuracy by up to 65%.
No amount of next-token prediction will reason about whether a tolerance stack-up in a turbine shroud interacts with thermal expansion under cyclic loading.
For teams building physical products, a few strategies have shown effectiveness.
Integration layer
Engineering data lives in STEP geometry, FMEA emails, GD&T annotations, and ERP BOMs; each format is a standalone integration challenge with no foundation-model API. Standards traceability spans ASME, ISO, and FAA Part 25 across decades of accumulated precedent. Field failure databases encode field data from programs that took years to run and cannot be regenerated from a training corpus.
Intelligence layer
Multi-agent orchestration that decomposes an engineering task into sub-problems, routes each to the right specialist, and reconciles their outputs into a coherent change recommendation, trade-off analysis, or compliance verdict with traceable citations. It provides a context graph over standards and failure-mode precedent, including ASME, ISO, and FAA Part 25 clauses, with versioning, applicability conditions, and supersession chains, linked to the failure investigations and field histories that cite them.
Physics solvers in the loop are invoked as tools, with meshes, boundary conditions, and results passed back into the reasoning chain. The numerical substrate that grounds proposals in what the physics actually permits.
Verification layer
But coupled to a verifier, the same models behave differently. RocketBench tested a 7B-parameter model with reinforcement learning on iterative rocket design. Grounded in a physics simulator loop, it outperformed GPT-4-class models and human experts. Without the simulator, all LLMs plateaued below human performance.
AlphaGeometry pairs a neural proposer with a symbolic deduction engine alone; the LLM solves almost none of the problems; together, they reach Olympiad-level performance.
AlphaProof couples an LLM with the Lean formal proof verifier and reached silver-medal level at the International Mathematical Olympiad. Yet in each case, the verifier was mostly a proof assistant.
By combining intelligence, integration, and verification
An engineer adjusting a clearance on a compressor blade tip sees that the change interacts with thermal expansion under cyclic loading, conflicts with the tolerance budget on two mating surfaces, and matches a failure signature from a prior program that led to a fleet-wide inspection.
The next generation of aircraft, vehicles, reactors, and robots are being designed by something better.