Development of hybrid machine learning based strategy for bioprocess online control

The background of this research is to explore and integrate a range of mechanistic and machine learning based modelling strategies to develop a multilevel digital framework (digital twin) for industrial bioprocess dynamic simulation and prediction, online optimisation (online learning), and product quality control. The framework will effectively rectify and covert low quality industrial plant data into a series of accurate predictive system models for process real-time digitalisation and state estimation. Cutting-edge online optimisation technologies will be also embedded into this framework to further support decision-making during an ongoing process.

In specific, this PhD project aims to resolve the challenges arising from bioprocess control, a major research area in the pharmaceutical and fermentation industry since the last decades. Despite the wide use of Extended Kalman Filter and Receding Horizon Controller in the chemical industry, a number of issues have been found when applying them into the biochemical industry due to the distinct characteristics of bioprocesses (e.g. high system nonlinearity, broad operational range, low reproducibility). Given the large number of data sets (e.g. high frequency online spectra data, low frequency offline numerical data) accumulated from real industrial plants, machine learning techniques may serve as a powerful alternative to resolve this task. This project will investigate how to combine unsupervised learning algorithms including principal component analysis, multiway partial least squares, and non-negative matrix factorization and supervised learning methods including different types of neural networks and Gaussian processes to develop an effective controller and software sensor for process real-time control and uncertainty estimation. Several essential procedures such as data dimension reduction, nonlinearity removal, and filling missing record will be systematically studied to maximise the performance of these hybrid machine learning based process control techniques. The possibility to extend these techniques into real-time process optimisation will be also explored and verified through collaborations (experimental research groups available from the UK and China). A brief introduction to the research group can be found through the following link:

https://www.research.manchester.ac.uk/portal/en/researchers/dongda-zhang(be2938f4-0e27-4c2c-9fcb-85ef46a1955d)/projects.html?period=running

Candidates are expected to hold or achieve a first class or 2:1 honours degree (or equivalent) in Chemical Engineering, Process Systems Engineering, Biochemical Engineering, Industrial Engineering, Mathematics, Computer Science or other related area. Students with a solid mathematical background are particularly welcome. A prior knowledge/experience in mathematical programming and optimisation theory is desirable.