Start date: Between 1st August 2027 and 1st July 2028
Application deadline: 23:59 on 4th January 2027 (UK time)
Introduction: This PhD addresses a central bottleneck in engineering design: generating a candidate geometry is fast, but evaluating its physical performance is not. Neural operators can predict multiphysics fields in milliseconds once trained, and generative models can propose new geometries in seconds. Yet these two capabilities are rarely combined into a single, closed-loop system. Doing so would let a designer generate, evaluate and refine a component in real time, rather than waiting on costly simulation cycles. This has direct relevance to thermal management, electromagnetic components and multifunctional lightweight parts.
You will train neural-operator surrogates — Fourier Neural Operator and DeepONet-style architectures — on thermal and electromagnetic performance data, for example, from multifunctional metastructures. You will then embed these operators directly inside a conditional diffusion or transformer-based generator, so physics feedback becomes differentiable and near-instantaneous during generation, rather than computed afterwards. You will embed governing equation constraints as physics losses within the generative architecture, develop conditioning schemes letting users specify target performance directly, and validate designs computationally. You will quantify the trade-off between operator fidelity and generative design quality. You will test whether a single joint multiphysics operator (thermal and electromagnetic together) can replace two separate surrogates without losing accuracy. You will build towards an interactive design-exploration tool with sub-second feedback, benchmarking throughout against conventional solver-in-the-loop generative pipelines to prove where your closed-loop approach genuinely wins. You will work across both computational validation and, where feasible, experimental testing via 3D-printed prototypes, keeping the work tied to physically realisable structures rather than a purely theoretical exercise.
You will be supervised by Dr. Ajit Panesar and Prof. Rob Hewson, whose complementary expertise spans generative design, engineering optimisation, structural metamaterials and scientific ML, and you'll be working alongside a talented group of researchers in a dynamic, intellectually stimulating environment.
Supervisors: Dr. Ajit Panesar and Prof. Rob Hewson
Duration: 3.5 years.
Funding: Full coverage of tuition fees and an annual tax-free stipend of £23,805 for Home, EU and International students. Find out more about fee status
Eligibility: Due to the competitive nature of our studentships, candidates will be expected to achieve/have achieved a First class honours MEng/MSci or higher degree (or international equivalent) in: Aeronautics or Mechanical Engineering
How to apply: Visit Generative Neural-Operator Surrogates for Real-Time Multiphysics Design for important instructions on how to apply.
Contact: For more specific project information, email: Dr. Ajit Panesar
For questions about the application process, please first read our Frequently Asked Questions. If your question has not been answered, please email the PhD Administrator (Admissions)
Frequently Asked Questions: You can also find answers to common application questions on our Frequently Asked Questions.
Equality, Diversity and Inclusion: Imperial is committed to equality and valuing diversity. We are an Athena SWAN Silver Award winner, a Stonewall Diversity Champion, a Disability Confident Employer and are working in partnership with GIRES to promote respect for trans people.
