Prediction of penicillin concentration using time series forecasting and machine learning techniques

Area of Science:

• Pharmaceutical Manufacturing
• Bioprocess Engineering
• Machine Learning

Background:

• Time-series forecasting is vital for pharmaceutical fermentation quality control.
• The industry's shift towards Pharma 4.0 necessitates advanced monitoring and control.
• A gap exists in reliable machine learning tools for complex pharmaceutical fermentations.

Purpose of the Study:

• To develop a Gated Recurrent Unit (GRU) based machine learning model.
• Predict penicillin concentration in fermentation processes.
• Support pharmaceutical manufacturing optimization and control.

Main Methods:

• Trained a GRU neural network on simulated penicillin fermentation data (IndPenSim dataset).
• Employed Random Forest and XGBoost for input parameter selection.
• Conducted systematic hyperparameter tuning for model optimization.

Main Results:

• Achieved a Mean Squared Error (MSE) of 7.20 × 10^-4 with the optimized GRU model.
• Data-driven feature selection enhanced predictive accuracy.
• Model demonstrated reliability in forecasting penicillin concentration.

Conclusions:

• GRU networks are feasible for penicillin concentration forecasting in fermentation.
• The GRU model serves as a practical tool for real-time process control and deviation management.
• This approach supports Pharma 4.0 by enhancing transparency and enabling digital twins in regulated environments.