Development of Transfer Learning techniques for the scale-up of biopharmaceutical processes

The aim of this Thesis is to evaluate whether a multivariate multi-block latent variable regression model, JYPLS (García-Muñoz et al., 2005), is as an effective tool to predict and optimize product quality through transfer learning from a pilot-scale plant to an industrial-scale one, ascertaining, at the same time, what is the appropriate number of pilot-scale batches required to transfer information from the pilot scale to the industrial scale. The case study considered in this Thesis is a simulated fed-batch process for penicillin fermentation. In particular, Pensim (Birol et al., 2002) is used to simulate the pilot scale, while Indpensim (Goldrick et al., 2014) is used for the industrial scale. The specific objectives are: i) to obtain the most accurate and precise estimations of the final penicillin concentration from process variables collected online, and ii) to optimize the product quality achieving the highest penicillin concentration possible at the end of the batch in the industrial scale through model inversion. Two methodologies for designing the experimental campaign for data collection in the pilot-scale plant are also evaluated: a full factorial design and happenstance design. It has been demonstrated that JYPLS is an effective transfer learning model, particularly in the case of well-controlled processes where random disturbances on the process variables are limited, even if low amount of data are available from a low number of batches (starting from 2 pilot-scale batches the model can be improved). Model accuracy and precision are improved by 5%, and productivity increased by 0.03%, which is equivalent to more than 1 kg of penicillin per batch.