Abstract
Advances in monitoring and computation have accelerated bioprocess digitalization, thus increasing the relevance of hybrid semi-parametric models. However, Machine learning tools struggle with collinearity in bioprocess datasets and lack of interpretability, where the selection of the biological and operational signals used for training the non-parametric component is still not straight forward. To address this, a three-step pipeline for feature engineering (i.e. SHAP values, mutual information, and domain knowledge) is proposed. Using this approach, hybrid models were built for ABE fermentation by Clostridium saccharoperbutylacetonicum, combining mechanistic mass balances and stablished kinetics with a Gaussian Process (GP) model for predicting butanol formation. Four datasets (three for training, one for validation) covering seven states were used, and 8,000 models were trained per strategy, enabling robust ensemble comparisons. SHAP ranked glucose, time, acetate, and acetone as the most influential features for the GPs. However, including time did not improve predictions, as mutual information showed its predictive content is largely redundant with metabolites. Conversely, initial glucose—ranked low by SHAP—was useful for anchoring batch variability. Hybrid models using glucose, initial glucose, and acetate outperformed both the parametric baseline and SHAP-guided choices, reducing validation RMSE by up to 12.8%. This strategy showed that integrating interpretability, redundancy analysis, and mechanistic insight yields more accurate and meaningful hybrid models than relying on importance scores alone.
