developing and validating advanced Bayesian optimization algorithms

This PhD project is about developing and validating advanced Bayesian optimization algorithms to drive the search for biopharmaceutical manufacturing processes that are more efficient in terms of costs, environmental impact, floorspace usage, etc. This means that drugs will be more accessible and cheaper for patients, while affecting the environment less.
The industry partner of this project, Biopharm Services, has developed a market leading computational tool called BioSolve to support drug manufacturers and suppliers in their quest to design more economical and sustainable production processes. BioSolve comprises a detailed Excel-based process economics-mass balance model that can be configured to create digital twins of different biopharmaceutical production processes. An initial Bayesian Optimizer (BO) has been connected to BioSolve, and the goal of this PhD project is to bring this optimizer to a new level.
Biopharmaceuticals are pharmaceutical drug products manufactured in, extracted from, or semisynthesized from biological sources. They include, for example, vaccines, gene therapies, and recombinant therapeutic proteins. Although very effective and accurate, important issues of concern are the cost of production of biopharmaceuticals (e.g. expensive resources, need for high purity), the environmental impact of some of the emerging production technologies (e.g. disposables), and the space needed for the production sites.
The focus of the project is to advance the BO's capabilities in terms at least one of the following problem properties:
1. Quantifying, simulating and accounting for the inherent noise in a drug production process.
2. Tackling problems or larger scale in the decision and/or objective space.
3. Allowing users to perform interactive multi-objective optimization.
All of these three capabilities will require methodological contributions, which can then be validated on real use cases provided by the industry partner (using also the BioSolve simulator) or on synthetic test problems from the global optimization literature (which may need to be modified to account for the various problem features). The research questions to be tackled in the scope of the project are:
1. Can we develop a (heterogeneous) noise-handling method (e.g. based on state-of-the-art uncertainty quantification methods) for BO that is able to deal with different types and degrees of noise across the multiple objectives?
2. Can we develop a Bayesian multi-objective optimizer (e.g. based on decomposition-based methods) able to solve problems with many mixed-type decision variables (between 20-40 variables of varying type) and many objectives (more than 2)?
3. How do we combine the methods developed in 1) and 2), and use them within an interactive optimization framework to guide the search towards a user-preferred part of the search space only?
4. What existing synthetic test problems can be used / how do we combine existing synthetic test problems to test the different capabilities of the developed BO?
5. Does the developed BO perform robustly on real biopharmaceutical production case studies provided by the industry partner (with BioSolve being used as the evaluation engine)?
These are just some examples - the successful student would be able to develop the research programme within the scope of the broad topic area.