122: Machine Learning for Predicting Microbial Survival Curves With Come-Up Time During Thermal Processing

Introduction

Ensuring food safety and quality during thermal processing relies heavily on accurate microbial survival models. Traditional models often suffer from limitations such as data dependency, model complexity or limited flexibility. Machine Learning (ML), known for its ability to handle complex and nonlinear relationships, offers a promising alternative. This study aimed to address these limitations by leveraging the power of Machine Learning (ML) to predict microbial survival curves. Specifically, we investigated the impact of come-up time (CUT) – the time required to reach the target processing temperature – on microbial inactivation.

Methods

Various machine learning models, including Support Vector Regression (SVR), Gaussian Process Regression (GPR), Artificial Neural Networks (ANNs), and Recurrent Neural Networks (RNNs), were used to fit microbial survival curves generated by the Weibull model with tailing. To improve accuracy, Adaboost was integrated as an ensemble method. Additionally, a novel numerical algorithm was developed to transform survival data with CUT to equivalent data under ideal constant conditions (CUT = 0), which served as input for training the ML models.

Results

Kernel-based ML models demonstrated excellent fitting performance, with GPR and SVR using the RBF kernel consistently outperforming other configurations. When these models underperformed, the incorporation of Adaboost significantly improved results, reducing Mean Squared Error (MSE) by up to 400%. For Neural Networks, the limited dataset posed challenges, as ANNs and RNNs exhibited overfitting tendencies. This was mitigated through careful architectural design, pre-training, and fine-tuning strategies. Additionally, the numerical algorithm developed in this work effectively estimates changes of the microbial survival ratio over time during non-isothermal heat treatments.

Significance

ML approaches excel at modeling complex, nonlinear microbial survival dynamics, enabling a single model to accommodate diverse patterns of raw data—an advantage over traditional mathematical models restricted to specific scenarios. Moreover, ML provides more accurate predictions, presenting a robust and versatile framework for enhancing the precision and efficiency of microbial inactivation modeling under varying thermal processing conditions.

Authors: Bing Li, Si Zhu, Guibing Chen