A hybrid mechanistic-data-driven framework for knowledge-driven and experiment-efficient HPLC method development

The traditional procedure for high-performance liquid chromatography (HPLC) method development requires many experiments to be performed, so it costs heavily on the experimental budget. To reduce the experimental budget, one can use a first-principle mechanistic model that, when accurately calibrated, can predict the separation process in-silico. Sometimes, due to limited understanding of the actual separation process, a mechanistic model cannot be calibrated to a satisfactory accuracy. This is when a data-driven model learning from data from past experiments can be used. Therefore, this work proposes a hybrid framework incorporating both mechanistic and data-driven models for knowledge-driven and experiment-efficient optimization. The proposed framework consists of two parts: (1) a mechanistic step where the HPLC method is optimized using a mechanistic model that is calibrated simultaneously; and (2) a data-driven step that is trained with the experiments from the mechanistic step. The results from two in-silico case studies demonstrate that, if the mechanistic model was well-calibrated in the mechanistic step, it was possible to obtain an optimal HPLC method for the cases considered using only the mechanistic step and with a total of eight experiments. Otherwise, if the mechanistic step fails to produce satisfactory results (e.g., ill-calibrated mechanistic model), the data-driven step then utilizes experimental information from the mechanistic step to generate an optimal solution. The performance of the proposed framework is compared to a pure data-driven approach, revealing that the hybrid approach is more efficient.