The rational design of single-site metal-doped amorphous materials is a grand challenge in heterogeneous catalysis. A brute-force approach for the description of these catalysts is not feasible because it would require millions of ab initio calculations of structurally different active sites to estimate the effective rate constant. This research will address this problem by developing algorithms and software that combine smart sampling and machine learning to systematically and efficiently focus the computational effort on active sites with unusually low activation energies that, as a consequence, dominate the catalyst kinetics. These algorithms will develop training sets on-the-fly while also using them to guide the sampling. The ab initio calculations necessary for the training set will be performed efficiently in parallel on thousands of cores, which is in line with the Exascale Computing Initiative of the Computational Chemical Sciences program. These calculations will be complemented with classical and ab initio molecular dynamics (MD) simulations, to explore the effect of the solvent on the structure and energetics of the active site. The results of the MD simulations will permit the refinement of the prediction of the site activity by repeating the machine learning analysis including explicit solvation. The software will be tested on two important reactions of ethylene that are catalyzed by metal-doped amorphous silicates: epoxidation and polymerization. The mechanism of the first reaction is known, and it will be used as a reference, while the mechanism for the second reaction is only partially understood, and it will be used to test the predictive power of the methods.
