Modeling microbial inactivation by high-pressure homogenization with a machine learning approach

This study leverages machine learning to create advanced predictive models for microbial inactivation during high-pressure homogenization (HPH). Unlike conventional models, which often focus solely on operating conditions, these models integrate additional factors, such as homogenizer-specific hydrodynamics, liquid media properties, and microorganism-specific characteristics. These factors are typically omitted in conventional models due to their wide variability across studies and the challenge of transforming them into a limited set of quantifiable variables. For instance, the influence of variations in homogenization valve geometry or changes in fluid viscosity are rarely incorporated, despite their significant impact on HPH outcomes. Through a comprehensive meta-analysis of literature data and the incorporation of dimensionless number to cluster diverse independent operating variables, various models, including artificial neural network (ANN) and random forest (RF), are trained and tested. While RF models exhibit faster runtimes without sacrificing performance compared to neural networks, a hybrid model was also devised to enhance prediction accuracy. This hybrid approach integrates RFs with the empirical Weibull model, linking microbial inactivation with applied pressure and the number of HPH passes. Notably, the hybrid model outperforms others, aligning well with expected inactivation trends. Challenges persist, such as the need for additional data and the inclusion of more relevant variables, underscoring the study's significance in advancing our comprehension of HPH's impact on microbial inactivation, thereby bolstering food safety and prolonging shelf-life.