AutoML-Driven Soft Sensors for Real-Time Monitoring of Amino Acids in Mammalian Perfusion Cultures

Mammalian perfusion culture offers high productivity for complex biologics but presents operational challenges in maintaining process stability. Real-time monitoring is essential to address such challenges while traditional analytical methods often fall short due to several limitations such as detection sensitivity and spectral overlap. Alternatively, data-driven soft sensors have gained traction for estimating key process variables indirectly. This study focuses on developing soft sensors for amino acids, key nutrient metabolites in perfusion processes, using automated machine learning (AutoML) approach. By leveraging daily measurements, the AutoML framework optimizes feature engineering, model selection, and hyperparameters to build accurate soft sensors with minimal expert intervention. Performance was further improved by tuning time budget parameter, and incorporating supporting amino acid measurements, particularly for low-performing sensors. The results demonstrate that AutoML effectively streamlines soft sensor development and enables real-time amino acid monitoring, thus paving the way for realization of digital twins in advanced biomanufacturing.