Navigating through the Maze of Homogeneous Catalyst Design with Machine Learning

Microkinetic models of homogeneous catalytic reactions are constructed using ab initio simulations to gain insight into mechanisms, rationalize catalyst performance, and inspire design. Using a data-driven approach, the classical linear free-energy relationships extended multiple regression study both reactivity selectivity catalysts build important interactions. Volcano plots demonstrated as general analysis framework when comparing potential energy surfaces related extract decisive parameters visualize breaks in mechanisms. Nonlinear including random forests, support-vector machines, Gaussian processes, artificial neural networks applied predict reaction yields, enantioselectivity, activation energies based on experimental computational data. The ability forge difficult chemical bonds through catalysis has transformed society all fronts, from feeding ever-growing population increasing life expectancies synthesis new drugs. However, developing systems is tedious task that requires tremendous discovery optimization efforts. Over past decade, advances machine learning (ML) have revolutionized whole way approach data-intensive problems, many these developments started enter chemistry. Meanwhile, similar field only their infancy. In this perspective, we outline our vision for future design role ML navigating maze. network connected neurons used computation. input processed with specific mathematical operations output transmitted neurons. propagated entire provide final strategy combining techniques data infer subsequent evaluations improve an objective. generate For molecules, they structures, sometimes together estimated properties. method natural selection sufficiently good solutions complicated search problems. Specifically, algorithm uses proposed solutions, which genetic such selection, mutation, crossover until satisfactory. inverting traditional process devising molecule or structure then evaluating its property; is, inverse property structures fulfilling property. representations capturing features can be compared between molecules allowing substructure searches, similarity cluster analyses, classifications. description graph theory. it represented vertices labeled element nodes bond type. area investigating agent take actions environment maximize total reward. Deep reinforcement deep problem. 100% robust string representation molecular formal grammar atomic valence rules. A set allowed SELFIES characters always corresponds molecule. simple SMILES does not necessarily correspond