Methodology for contamination detection and reduction in fermentation processes using machine learning.

Area of Science:

• Biotechnology
• Machine Learning
• Industrial Microbiology

Background:

• Fermentation processes are susceptible to contamination, impacting yield and product quality.
• Accurate and timely detection of contamination is crucial for industrial bioprocesses.
• Traditional contamination detection methods can be slow and lack sensitivity.

Purpose of the Study:

• To develop and validate accurate and efficient machine learning (ML) methodologies for fermentation contamination detection.
• To optimize hyperparameter tuning for ML models using advanced algorithms.
• To compare ML-based detection with traditional threshold-based methods and identify key factors influencing contamination.

Main Methods:

• Utilized two ML methods: one-class support vector machine (OC-SVM) and autoencoders.
• Employed Optuna platform for parallel hyperparameter optimization (HPO), recommending Bayesian optimization with the Hyperband algorithm.
• Benchmarked ML models against the traditional mean ± 3σ rule for contamination detection.

Main Results:

• Achieved high recall (up to 1.0) for detecting contaminated fermentation batches without compromising precision (up to 0.96) and specificity (up to 0.99) for non-contaminated batches.
• OC-SVM demonstrated superior precision and specificity compared to autoencoders, with both methods achieving perfect recall.
• Identified key independent variables contributing to contamination, providing actionable insights for process regulation.

Conclusions:

• ML models offer a highly accurate, data-driven approach for real-time fermentation contamination detection.
• These models require minimal retraining and are suitable for integration into industrial monitoring systems.
• The developed methodology effectively reduces contamination likelihood and surpasses traditional detection techniques.