TLDR: Fully-funded PhD scholarship on the application of transfer learning to Bayesian Optimization in the context of Computational Fluid Dynamics
Supervisors: Saleh Rezaeiravesh and Manuel López-Ibáñez (University of Manchester)
Application deadline: Friday, December 5, 2025
Start date (if accepted): September 2026
Apply here: https://www.ai-decisions-cdt.ac.uk/apply/
For any questions please contact the UKRI AI Decisions CDT Team (aidecisionscdt@manchester.ac.uk)
In many fields, including CFD (computational fluid dynamics) problems, the use of Bayesian Optimisation (BO) leads to faster experimentation, prototyping and analysis.
Bayesian Optimisers are data-driven methods that incrementally build surrogate models. However, building accurate surrogate models often requires expensive simulations and laboratory experiments. Even small changes of the conditions under which the data was collected may alter the landscape modeled by the surrogate, which would require collecting new data and rebuilding the surrogate from scratch. If building the original surrogate was very costly and the new landscape is closely related to the previous one, throwing it away is a net loss. In such scenarios, transferring the original model to the new landscape may significantly reduce the amount of new data that needs to be collected to obtain an acceptably accurate model. We have recently proposed using a small sample of new data to transfer already trained surrogate models to a new landscape by optimising affine and nonlinear domain warping transformations.
The goal of this project is to further expand the capabilities of the transfer learning by researching more flexible and more efficient transfer techniques in the domain of CFD, where each new data point may require a costly simulation or real-world experiment. Practical use will also require ML techniques that can intelligently decide when active learning may be beneficial and when the model cannot be transferred to the new setting. Finally, the project will also investigate how to extend these techniques to ensembles of trained models rather than one model at a time.
Y. Morita, S. Rezaeiravesh, N. Tabatabaei, R. Vinuesa, K. Fukagata, P. Schlatter, Applying Bayesian optimization with Gaussian process regression to computational fluid dynamics problems, Journal of Computational Physics, 2022. https://doi.org/10.1016/j.jcp.2021.110788
Shuaiqun Pan, Diederick Vermetten, Manuel López-Ibáñez, Thomas Bäck, and Hao Wang. 2025. Transfer Learning of Surrogate Models: Integrating Domain Warping and Affine Transformations. In Proceedings of the Genetic and Evolutionary Computation Conference Companion, Association for Computing Machinery. https://doi.org/10.1145/3712255.3734291