Capturing uncertainty in black-box chromatography modelling using conformal prediction and Gaussian processes

We demonstrate that conformal predictors – specifically conformalised quantile regression (CQR) and locally adaptive conformal predictors (LACP) – outperform the commonly used Gaussian Process Regression (GPR) in uncertainty quantification of machine learning surrogate models for chromatography modelling. CQR excelled in black-box scenarios, effectively estimating challenging target variable distributions, while LACP provided extremely informative intervals when kinetic parameters were included. Incorporating kinetic data significantly reduced epistemic uncertainty and increased model accuracy, supporting the hypothesis that adding mechanistic data to black-box models improves prediction uncertainty. This study represents the first application of conformal methods in chromatography modelling, indicating high applicability of this new uncertainty quantification methodology. Our findings offer a promising direction for advancing uncertainty quantification methods in data-driven bioprocess modelling and optimisation.