Jonathan Raines

General Profile:

I have an engineering background and have worked in industry for 9 years before starting this PhD. I worked at Rolls-Royce for a short time but have mainly been at a company called Open Bionics which makes robotic prosthetic hands for upper-limb amputees. I've had a small amount of contact with AI through the lens of robotics. I'm interested in how it can be used in Engineering by speeding up simulations, generating designs, or parsing complex regulations and standards. Additionally, as my background is in the medical device industry where risk management is at the heart of development, I'd like to explore applying those principles to AI safety.

Research Project Summary:

This PhD will create an interactive design assistant for engineers that can generate solutions that include library components.

Engineering Design can be seen as a search for a solution through a complex "design space". Time spent searching can yield better (more functional, more appreciated by users, more sustainably made) solutions. However, in industry, the time spent on design is constrained by cost and time-to-market requirements.

Generative Design is the application of algorithms to help search. Present Generative Design tools produce unique, free form, highly complex and intricate designs. When paired with 3D printing, extremely high-performance parts have been made but at high cost and low levels of quality assurance. For many industrial applications it is preferable to build (in part or in full) from standardised parts that can be bought in or manufactured quickly in bulk. The automotive industry is an example of this. There is a "tall" vertical supply chain. Manufacturers assemble from components purchased from top-level suppliers who do the same to lower-level suppliers. Additionally, manufacturers make various products and can lower costs by sharing parts between them.

The project will investigate two aspects:  (1) How can the inclusion of standard parts improve Generative Design outcomes/solutions in terms of performance, cost, and manufacturability? (2)How can the recent advances in Generative AI and Reinforcement Learning be applied to consider standard parts during Generative Design? Using these results, AI methods will be applied to enable the use of a library of standard parts. 
The first stage involves applying Computer Vision techniques to recognise portions of Generative Design parts that could be replaced with standard parts. This process mirrors a manual technique sometimes applied by engineers. The free form part is approximated by a combination of off-the-shelf parts. It can be thought of as a "find-and-replace" operation.

The second stage is building designs from scratch by selecting a combination of parts. Phrasing the problem like this turns it into a form of combinatorial optimisation. The challenge here is the number of possible configurations is large for even small libraries. Here, Reinforcement Learning has shown promise in domains with similar characteristics, such as the games of Chess and Go. Recent work has shown promise in applying this to engineering design. However, there are still challenges as engineering design problems are often multi-objective; they involve a mix of categorical and continuous variables, and quantifying how well a solution meets an objective frequently involves physical testing or computationally expensive simulation.

In addition to the results of the research assisting in searching the design space more efficiently, AI can also help reduce the computational cost of simulation via surrogate modelling. The cost of data collection and training a surrogate must be offset by the savings in running it. Therefore, researching generalisable, reusable surrogates is preferable.

Supervisors:

Website:

Edit this page