Tools in the loop, not a replacement for it
Artificial intelligence now appears in several parts of the design process, but it helps to be specific about what that means. In engineering design, “AI” usually refers to software that searches for, suggests or predicts something based on data or rules. It does not understand the product, the customer or the consequences of failure in the way an engineer does. The useful question is not whether AI will design for us, but which steps in the loop it can speed up and what checks it needs.
For the wider picture of how AI affects technical work, see the discussion of artificial intelligence as a field; this article stays with the design loop.
Where AI can fit
Generative design
Generative design tools take a description of a problem, such as where loads and supports are, how much material may be used and which regions must stay free, and then produce one or more candidate shapes. Internally, many of these tools rely on optimisation methods that adjust the geometry to improve a measurable goal, such as lower mass for a given stiffness.
The result is often an organic, unfamiliar-looking shape. Whether it is a good design depends on what the tool was told. Anything left out of the problem statement, such as how the part is made, how it is inspected or what happens when a load comes from an unexpected direction, is not considered.
Optimisation
Optimisation does not have to be exotic. Adjusting a handful of dimensions to reduce weight, cost or temperature while meeting limits is a classic task. Search methods can try many combinations faster than a person, and some use machine learning to decide which combinations to try next. The quality of the answer depends on the model that scores each trial and on the objective chosen. If the objective is incomplete, the optimiser may find a design that scores well and is poor in practice.
Surrogate models
Detailed analysis can be slow. A surrogate model is a faster approximation trained on results from earlier runs, and it can be used to explore options quickly. The usual caution applies: the surrogate is only reliable near the conditions it was trained on, and results should be confirmed with the full analysis, such as proper engineering simulation, before anything is committed.
Assistive tasks
AI is also used for smaller jobs: suggesting parts from a catalogue, finding similar earlier designs, flagging features that look hard to manufacture, summarising documents or helping draft text. These can save time, provided the output is read critically. Errors in an automatically generated summary or suggestion are easy to miss because the text sounds confident.
Honest limits
- Dependence on inputs. The software optimises the problem it is given. If requirements are incomplete or wrong, the results will be too.
- Dependence on data. Learning-based methods reflect their training data. Unusual conditions, new materials or new failure modes may lie outside what the tool has seen. A short grounding in how machine learning works makes this limitation easier to see.
- Opaque reasoning. Some tools cannot explain why they suggest a shape, which makes review harder.
- Manufacturing and cost. A shape that is mathematically efficient may be difficult or costly to produce, depending on the process. Check that the design suits the way you intend to build it.
- Models are approximations. Any analysis behind the suggestion has assumptions, and those assumptions need to be checked against reality.
- Accountability. Software does not carry professional responsibility. People do.
The engineer’s role
Engineer review is not a formality. Treat AI output as a proposal from a fast but uninformed colleague.
- Define the problem carefully. Write requirements, loads, constraints and the real goal before running anything.
- Question the inputs. Check units, load cases, material properties and boundary conditions.
- Inspect the result. Does it make physical sense? Does it resemble a known good solution? If it looks surprising, find out why before trusting it.
- Verify independently. Use hand calculations, separate analyses or tests to confirm key results. Order-of-magnitude estimates are valuable here.
- Consider the full life. Think about manufacture, inspection, repair, failure modes and end of life.
- Document the decision. Record what tool was used, what it was asked to do, and why the final design was accepted. Short decision records suit this well.
Design reviews remain a good place to challenge AI-assisted work, because other people ask questions the original user may not have considered.
Responsible use
Some practical points of care:
- Be cautious about uploading confidential designs to external services. Check your organisation’s rules and the tool’s terms.
- Check whether the licence or terms of use of a tool affect who owns or can use the output.
- In safety-related fields, regulators and standards bodies may set expectations for how software tools are qualified and how results are verified. These rules vary by country and industry, so follow those that apply.
- Do not let a polished result substitute for understanding. If you cannot explain why a design works, you are not ready to release it.
A cautious outlook
Tools in this area change quickly, and it is hard to predict which will prove lasting. A sensible stance is to learn the underlying ideas, such as objectives, constraints, optimisation and validation, since those outlast particular products. Use new tools on low-risk problems first, compare their output with methods you trust, and keep the human judgment at the centre of the design process.