Adapting AI isn't enough. Advanced manufacturing needs to build its own.
Insights from: Stephen McCartney - Chief Solutions Architect, Fairuz Yooseff - Technology Programme Manager at NCC
Almost every industry is looking at what AI means for their field.
Most conversations focus on using or adapting existing AI: embedding chatbots like Copilot into day-to-day operations, or fine-tuning ChatGPT Enterprise on internal data.
But advanced manufacturing is one place where off-the-shelf models don’t cut it. Existing models can’t manage a production process. They won't help wind turbine manufacturers deliver on time.
The processes we work on at NCC need bespoke AI, built from the ground up.
What building actually involves
AI is prediction. Large language models predict text because they've been trained on text. They're built to read and write, not to run a production process.
That’s because AI is like a toddler: show it one thing and it assumes the world works that way. It needs enough labelled examples to learn which decision is right.
Building a bespoke AI model for engineering and manufacturing takes five things:
-
Data scientists who can determine the right model for the process.
-
Domain experts, usually engineers, to translate shop floor know-how into something the model can learn from.
-
High-performance compute to build and test the model.
-
Enough labelled training data to learn from.
-
Physical testbeds to prove the model works.
NCC has all five under one roof.
Our work in AI-controlled liquid resin infusion shows this in practice. Sensors watch resin flowing around a mould. The model, trained on thousands of simulated runs, predicts if the injection will produce an air bubble. If it identifies a potential issue, the input pressure changes before the issue becomes critical. Result: fewer defective parts, less re-work, lower waste.
Because the model watches the flow rather than the kit, the same approach works across different processes and factories. Defect detection works the same way and is an area we’re actively exploring. Show a model enough images, scans, or ultrasound traces of defect-free parts and it reliably flags anything that isn't.
Bigger problems need bigger compute
Building bespoke models takes serious compute. We have access to Isambard-AI, the UK's national AI supercomputer, run by the University of Bristol on behalf of UKRI. It means we can iterate orders of magnitude faster and run techniques our own GPUs couldn’t handle. Problems too costly to run elsewhere become viable.
Manufacturing data is often scarce. When it is, we use Isambard-AI to run physics-based simulations to generate data. The resulting model isn't as accurate as one trained on real measurements, but it's close enough to fine-tune from there.
IT and Engineering join forces
None of this works without the right foundations. AI needs structured data and integrated systems. Engineering teams need infrastructure that keeps production running and supports the projects that bring new technology onto the shop floor. Partners need to know their IP is secure.
For years, IT teams at manufacturing companies were kept at arm's length from the shop floor. That worked when the two were separate. It doesn't now the shop floor is on the network. At NCC, our Digital Ways of Working programme brings those teams together, redefining how Engineering Technology, IT, and Operational Technology relate. The aim is simple: IT as enabler, not blocker.
Putting it into practise
Most manufacturers don't need to build all five capabilities themselves. They need a place to bring their problems, test what works and prove a model on real kit before committing. That's what NCC can provide for industry.
At NCC, we connect cutting-edge digital capability with the real-world challenges on the workshop floor.
Through deep links with industry, government and academia, we help turn new ideas into useful tools - and help guide businesses through the process.