Modern computer aided approaches to rationalize and predict olefin polymerization catalysis

Machine learning (ML) predictors are hailed as tools that could help faster development of better performing catalysts. This vision rests on the capability of ML-predictors to identify which catalyst structures outperform known ones. Nevertheless, several problems must be solved before ML-predictors can become a standard tool in the ordinary catalysis lab. In this joint presentation we will focus on some of these issues, and we will propose strategies to build robust ML-predictors. One critical point in developing ML approaches for catalyst engineering in homogenous catalysis is the scarce availability of large and curated training datasets. In most of the cases only few dozens of catalysts are available, and synthesis of new ones can require remarkable efforts. Unfortunately, it is difficult to gauge the accuracy of ML-predictors, trained on few-entry datasets, to identify improved catalysts before deployment in a real-world domain. To solve this issue, we developed a series of tests and confidence scores aimed to estimate the expected prediction capability of a new ML-predictor. To achieve robust validation of the workflow we considered a variety of different reaction classes, covering a broad chemistry scope. High-throughput experimentation databases offer a decisive advantage over mixed databases curated from literature sources as they ensure judicious adherence to standard process conditions. Utilizing such databases, we have developed catalyst class independent models for key performance parameters in olefin polymerization catalysis using a limited number of “chemically intuitive” descriptors that partition the distribution of steric bulk in the active pocket and measure electronic effects. Descriptors were designed for the performance indicators stereoselectivity and regioselectivity. The power of these new descriptors is emphasized by (a) models describing molar mass capability and comonomer affinity “out-of-the-box” and (b) the generated chemical insight.