Generative data augmentation and automated optimization of convolutional neural networks for process monitoring

Chemometric modeling for spectral data is considered a key technology in biopharmaceutical processing to realize real-time process control and release testing. Machine learning (ML) models have been shown to increase the accuracy of various spectral regression and classification tasks, remove challenging preprocessing steps for spectral data, and promise to improve the transferability of models when compared to commonly applied, linear methods. The training and optimization of ML models require large data sets which are not available in the context of biopharmaceutical processing. Generative methods to extend data sets with realistic in silico samples, so-called data augmentation, may provide the means to alleviate this challenge. In this study, we develop and implement a novel data augmentation method for generating in silico spectral data based on local estimation of pure component profiles for training convolutional neural network (CNN) models using four data sets. We simultaneously tune hyperparameters associated with data augmentation and the neural network architecture using Bayesian optimization. Finally, we compare the optimized CNN models with partial least-squares regression models (PLS) in terms of accuracy, robustness, and interpretability. The proposed data augmentation method is shown to produce highly realistic spectral data by adapting the estimates of the pure component profiles to the sampled concentration regimes. Augmenting CNNs with the in silico spectral data is shown to improve the prediction accuracy for the quantification of monoclonal antibody (mAb) size variants by up to 50% in comparison to single-response PLS models. Bayesian structure optimization suggests that multiple convolutional blocks are beneficial for model accuracy and enable transfer across different data sets. Model-agnostic feature importance methods and synthetic noise perturbation are used to directly compare the optimized CNNs with PLS models. This enables the identification of wavelength regions critical for model performance and suggests increased robustness against Gaussian white noise and wavelength shifts of the CNNs compared to the PLS models.