Hybrid modelling based digital technology for process flow diagram development

Smart manufacturing is one of the novel concepts arising from the era of the 4th Industrial Revolution. To remain competitive both regionally and globally, it is critical for companies to efficiently develop more sustainable and personalised products in order to achieve a significant increase in process profit, reduction in energy cost and waste generation, and promotion of customer satisfaction. However, the present industrial process flow diagram development approach usually undergoes several essential but time-consuming steps, ranging from lab-scale product formation to pilot-scale R&D process design, and eventually transformation to factory manufacturing, causing a long product development cycle.

Given the development of transformative artificial intelligence technology and the large amount of data accumulated in the process industry, combining machine learning based data-driven models with first-principle models such as computational fluid dynamics (CFD), as well as statistical analysis, offers a paradigm change in the way that product formulation and process development will be planned and executed in future manufacturing. A novel approach to predictive and hybrid digital modelling technology will be required to effectively estimate key properties for product formulation under different ingredients and to facilitate process design, operation, diagnosis and knowledge transfer across different manufacturing lines. This digital driven process flow diagram development approach will also allow sensitivity and uncertainty analyses to account for external factors such as raw material variability, data measurement noise, and site-to-site variation.

Building upon recent success and collaborative culture between the University of Manchester and Unilever, the main objective of this PhD project is to investigate state-of-the-art hybrid modelling and transfer learning strategies to accelerate process flow diagram development and automate process operation. Particularly, we will explore physics-informed hybrid models to maximise process data utilisation and combine rigorous physical models with high-fidelity machine learning techniques. The developed hybrid models will be applied to predict and optimally control performance of essential unit operations and identify key operational activities within personal care product manufacturing processes, and we will test the transferability and robustness of the developed models across different sites.