Applying AI and Simulation Outputs to Real Engineering Decisions

The hypothesis

My own synopsis said neural networks can approximate simulation outputs "without compromising accuracy". I wrote it, and then spent the talk testing it. It doesn't hold as a general claim. A surrogate inherits every error in the runs it learned from and adds its own. The question worth asking is whether it's accurate enough for the decision in front of you.

Four ways AI gets into a simulation workflow

Predict: surrogate models trained on past runs. Generate: new geometry from requirements or a prompt. Build: coding agents that write scripts and niche solvers. Drive: agents operating the software you already have. None of them replaces your solver.

Explore, Narrow, Commit

Explore. Generate lots of options, fast and rough. You want the right trends, not exact numbers, so this is where AI does the most work.

Narrow. Check the shortlist meets your requirements, check the surrogate behaves and spot-check it with the solver.

Commit. Back to full solver and physical test, run in parallel with the old process. The AI output stays out of the evidence.

How accurate it needs to be depends on the decision you're making with it.

What the published cases show

GM. A model trained on GM's own CFD data predicts drag inside the designers' sculpting tools. A two-week loop is now minutes. The figure used for fuel-economy certification still comes from a wind tunnel. (GM newsroom, April 2026; IEEE Spectrum, April 2026)

JLR. Aero evaluations went from about 50 to about 1,500 a day, using surrogates trained on more than 20,000 of JLR's own simulations and checked against CFD and the wind tunnel. (Vendor-reported by Neural Concept, from JLR's NVIDIA GTC talk, March 2026)

Dallara. Set two accuracy bars before training: one for concept work and a stricter one for standing in for CFD. The model cleared the first on all 20 parts and the second on 18. (arXiv 2604.18491, April 2026)

A linear FEA solver running in the browser

A browser FEA bench with a neural surrogate checked against the solver on every change. Inside its training range the surrogate is within 1%. Outside it, you can watch the error climb.

Six checks before a surrogate result counts

Behaviour: minimum functionality, invariance, directional expectation.

Decision: inside the training envelope, same ranking as the solver, obeys the physics.

Then verify the shortlist against your requirements.

The guide explains each check with worked simulation examples.

Nine failure modes in agentic simulation workflows

Objective drift, cost explosion, memory corruption, cascading errors, tool drift, prompt injection, infinite loops, context decay, instruction drift. Most trace back to three gaps: no contract, no checks between steps and no record of what the tool did.

Questions people ask

About the speaker

Ali Parandeh is a Chartered Engineer (IMechE) and founder of Build Your AI. He has spent 13 years across rail signalling, rail AI and generative AI consultancy, and is the author of Building Generative AI Services with FastAPI (O'Reilly) and the Building Reliable AI Agents liveProject series (Manning).

If you're working out how far to trust AI on a live project, get in touch.

Sources