PREDICTIVE MODELS FOR CATALYST DEVELOPMENT

Abstract

With these hard times that we are living after covid, inflation but also problems like fertilizer shortage and supply of chain issues, has made everyone turn their attention to better, more affordable, faster, and organic solution almost in every field of science and not only. The inspiration for this project was found on the BioSPRINT project, where the target reaction is the simultaneous dehydration of multiple C5 and C6 sugars to produce 5-HMF and FUR. The objective was to find machine learning (ML) models that would speed up the discovery of catalysts using high-throughput (HTP) screening techniques. Maximum activity for the conversion of complex sugar combinations is sought, with the best selectivity for the major products of interest. The three additional models used are generalised boosted regression modelling, extreme gradient boosting and boosted generalised additive models for location, scale, and shape. The results show that XGBoost has the best performance overall. All the models performed poorly in the case of Selectivity. Another approach for this response is to apply a transformation on the response variable. The performance of these models can be potentially improved by adding new “catalytic-informed” features, that will be engineered based on the expert knowledge about the problem.