Soft sensors based on artificial neural networks for advanced bioprocess monitoring

Abstract This study presents the development and validation of artificial neural network (ANN)‐based soft sensors for the online estimation of key bioprocess variables during fed‐batch cultivations of Pseudomonas putida KT2442 producing medium‐chain‐length polyhydroxyalkanoates (PHA), sustainable biopolymer precursors for bioplastics production. Since variables such as cell dry weight (CDW) and PHA concentration are typically accessible only through offline laboratory analyses, soft sensors represent a promising tool for implementing advanced process analytical technology (PAT) strategies for monitoring and control. Historical experimental data from two nominal cultivations and two cultivations with non‐standard behaviour, including significant feeding perturbations, were used for model development. ANN‐based soft sensors were designed in both global and multi‐phase configurations, where dedicated models were assigned to individual cultivation phases. Model training and validation were performed using leave‐one‐batch‐out (LOBO) cross‐validation to assess robustness and generalization capability. The best‐performing soft sensors achieved normalized root‐mean‐squared errors of approximately 4% for both CDW and PHA concentration estimation. The highest prediction accuracy was obtained for models utilizing capacitance measurements as input variables, followed by models combining capacitance measurements with cumulative oxygen consumption and carbon dioxide production data. CDW soft sensors demonstrated higher robustness, maintaining reliable performance even for cultivations with non‐standard trajectories, whereas PHA soft sensors showed substantially higher sensitivity to process deviations and perturbations.