A Data-efficient Transfer Learning Approach for New Reaction System Predictive Modelling

Abstract

Transfer learning provides an effective and practical solution to modelling novel systems when a lack of theoretical understanding and data availability hinders progress. In this chapter, transfer learning aims to leverage previously discovered relations and prior understanding of complex biochemical systems to support the rapid construction of accurate predictive models for different but related biochemical systems. This chapter explores the application and advantages of transfer learning for a real experimental case study to demonstrate the potential of transfer learning within the biochemical industry. To maximise the use of available process knowledge, transfer learning and hybrid modelling are combined for the first time. Building on the hybrid modelling methodology introduced in Chapter 3, a step-by-step explanation is provided for transfer-hybrid model construction, focusing on the selection and implementation of the chosen transfer learning approach and the decision about which aspects of the model to transfer or update for the new system to avoid inheriting domain-specific biases. The study concludes by comparing the accuracy and uncertainty of the transfer-hybrid model with a traditional-hybrid model. Although the results are case-specific, they provide valuable evidence that transfer learning can accelerate biochemical process model construction and help bolster innovation when correctly employed.