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

BACKGROUND: Time-series forecasting is crucial in pharmaceutical fermentation, where early detection of deviations safeguards product quality. As the industry evolves toward Pharma 4.0, integrating machine learning into process monitoring and control presents an opportunity to enhance efficiency, consistency, and regulatory compliance. However, currently there is a notable lack of reliable machine learning based predictive tools for complex pharmaceutical fermentations.

OBJECTIVE: To develop a machine learning model based on Gated Recurrent Units (GRUs) for predicting penicillin concentration within a fermentation process, supporting process optimization and control in pharmaceutical manufacturing.

SIGNIFICANCE: Effective time-series forecasting of penicillin production is crucial for proactive control of critical quality attributes (CQAs). Integrating artificial intelligence techniques aligns with the goals of Pharma 4.0, enabling data-driven decision-making. This study demonstrates how a GRU-based soft sensor can support Pharma 4.0 by enabling accurate short-term forecasts of penicillin concentration under realistic industrial conditions. Using the IndPenSim dataset, which includes batch-to-batch variability and process disturbances, the model provides transparent, regulator-friendly predictions through integrated feature importance analysis. In practice, the GRU can act as an early-warning tool, alerting operators to deviations and allowing proactive adjustment of input variables.

METHODS: A GRU-based neural network was trained on a simulated dataset of 100 penicillin fermentation batches from the IndPenSim platform. The model development procedure included input parameter selection using Random Forest and Extreme Gradient Boosting (XGBoost), followed by hyperparameter tuning across various activation functions, optimizers, and dropout rates.

RESULTS: The GRU model trained with the Adamax optimizer and tanh activation achieved a Mean Squared Error (MSE) of 7.20 × 10-4. The data-driven feature selection approach improved predictive accuracy over manual selection. Graphical comparisons between the predicted and actual concentrations supported the model's reliability.

CONCLUSIONS: This study shows the feasibility of GRU-based neural networks in forecasting penicillin concentration during fermentation. Beyond theoretical evaluation, the GRU framework provides a practical industrial use case: real-time forecasting of penicillin concentration to support proactive process control and deviation management. Such applications directly contribute to Pharma 4.0 initiatives by enabling digital twins, enhancing process transparency, and lowering barriers for deployment in regulated environments. Compared to recent approaches, our method emphasizes systematic hyperparameter tuning and input feature interpretability while maintaining computational efficiency. Our GRU model achieves competitive accuracy with a simpler, more transparent architecture, highlighting its suitability for real-time pharmaceutical deployment in regulated Pharma 4.0 environments.