Abstract
Biopharmaceutical manufacturing requires systematic identification, understanding and control of critical process parameters (CPPs) and critical quality attributes (CQAs) While deterministic machine learning models have proven valuable for describing complex systems and providing insights into process behaviour, the extension to probabilistic frameworks allows for the capture of intrinsic biological and process variability, improving robustness, understanding and safety. This work introduces a framework for autoregressive variational Bayesian neural networks (BNNs) that integrate uncertainty quantification into data-driven modelling. The approach is demonstrated on a multicolumn countercurrent solvent gradient purification (MCSGP) process for monoclonal antibody purification. The BNN is trained autoregressively on high fidelity data generated using a detailed mechanistic process model data. A heteroscedastic variance head is included to model signal-dependent observation noise, while the epistemic uncertainty is captured through stochastic network parameters. Model performance is assessed using both deterministic accuracy metrics -MSE and probabilistic scores namely coverage and continuous ranked probability scores (CRPS). Results show that the proposed method accurately recovers noiseless concentration profiles while maintaining well calibrated predictive distributions. Additionally, the framework is tested under multiple noise regimes to test robustness of the proposed approach.