Product Attribute Forecast: Adaptive Model Selection Using Real-Time Machine Learning

Abstract

A real-time machine learning framework is developed to forecast product concentration in mammalian cell culture bioreactors. In real-time, the framework evaluates several machine learning algorithms and chooses the most representative algorithm based on current dynamics of the system. Data from multiple sources is combined and only subset of features are fed to the model based on a pre-selection criteria. The model performance is tested using two small-scale bioreactors run. The performance improved towards the end of the process with accumulating data and results for 1 day ahead prediction is presented.