End-to-End High-Throughput Approach for Data-Driven Internal Donor Development in Heterogeneous Ziegler–Natta Propylene Polymerization

Abstract

Internal donors (IDs) play a decisive role in shaping the structure and performance of Ziegler−Natta catalyst formulations for isotactic polypropylene production. Unfortunately, their diverse and intricate functions remain elusive, and rational ID discovery, therefore, is still problematic. Exploitation of artificial intelligence methods such as machine learning, in turn, has been hindered by the lack of training data sets with adequate quality and size. This study proposes an integrated high-throughput workflow encompassing catalyst synthesis, propylene polymerization, and polypropylene characterization. Its application to an ID library of 35 molecules generated a robust and consistent data set, which highlighted important and intriguing quantitative structure− property relations (QSPRs). Furthermore, by fingerprinting ID molecular structure in combination with feature selection, a black box QSPR model correlating ID molecular structure and catalytic performance was successfully implemented. This study demonstrates that the combination of high-throughput experimentation and machine learning is a promising asset for accelerating the research and development of Ziegler−Natta catalysts.

From Characterization to Discovery: Artificial Intelligence, Machine Learning and High-Throughput Experiments for Heterogeneous Catalyst Design

“…Automatic feature engineering: Used to increase the number of features of small data sets of multielement catalysts . It consists of the inclusion of elemental and molecular characteristics of the elements and molecules taken from all data sources possible (feature assignment).…”

“…Taniike et al (2024) used a combination of HTE-generated databases and GAs for the study of descriptors of Ziegler–Natta catalyst formulations for isotactic polypropylene production, with a focus on the roles of internal donors (IDs). An ID library of 35 molecules was methodically screened using an integrated HTE workflow, generating a data set that unveils quantitative structure-properties relations.…”

“…Some ML algorithms are particularly suitable for small data sets: SVM, GPR, RF, gradient boosting, XGBoost, and symbolic regression . Some strategies to deal with the problem of small data sets or experiments that are expensive to conduct and therefore require as little experimental data generation as possible include: Automatic feature engineering: Used to increase the number of features of small data sets of multielement catalysts . It consists of the inclusion of elemental and molecular characteristics of the elements and molecules taken from all data sources possible (feature assignment).…”