A generic experimental data acquisition approach to predict optimal combinations of analytical separations in multidimensional separations

The starting hypothesis for this PhD is that the principles underlying the Predictive Elution Window Stretching and Shifting (PEWS2) method developed by Tyteca [1] for method optimization can, either in its original or an adapted form, also be used for method selection. After some initial scouting runs, the PEWS2-approach applies a minimalistic degree of modelling to predict how well the first and last part of a chromatogram can be stretched out along a given chromatographic dimension. It is assumed that, by repeating this in different dimensions, we will be able to rapidly screen through the total separation space (=search space) of any potential multi-dimensional separation and find the combination of dimensions providing the largest separation capacity. This hypothesis will be investigated via two parallel routes. One route will be purely in silico, studying model-based representations of randomly composed mixtures of molecules, each characterized by a set of randomly attributed descriptors defining their elution characteristics in all possible elution modes. Since the composition of these in silico samples will be exactly known, the computer will know the best (combination of) separation mode(s) beforehand, allowing to unambiguously verify whether the proposed experimental search algorithm can successfully guide the search to this solution or not. To scout through the vast number of combinations, machine learning techniques such as deep learning network architectures and random forest models will be used. In a later stage, we will move to in silico samples where components are grouped in sample dimensionality groups. The other route will be experimental, where the PEWS2- based searches will be applied to a number of complex separation problems from different application fields (environmental, (bio)pharmaceutical, and biomass analysis). In a later stage, we will further extend the search space, and hence the separation space, by allowing for “sample splitting” routes, by translating the sample splitting ideas initially proposed by Cabooter [2] for method development to the level of method selection. [1] Tyteca et al., JCA 1368 (2014) 125 [2] Loos et al., JCA 1497 (2017) 70