Automated Predictions of Chemical Reactions and Their Mechanisms

K.K. Irikura and R.D. Johnson III

Objective: To predict the reactivity of molecules.

Problem: Chemical reactions are of essential and fundamental importance throughout chemistry and related technologies. Although experienced chemists can sometimes predict the reactions that will occur in a new chemical system, they may overlook some alternatives. Moreover, they are usually unable to make reliable predictions when the chemistry in question is unfamiliar to them. As more exotic chemicals and materials are investigated, this situation is increasingly common. Yet there are few tools available to assist in predicting chemical reactions, and none at all for predicting the novel reactions that are of greatest interest.

Approach: Quantum chemical calculations can predict how the energy of a chemical system changes as its constituent atoms move. This energy function, known as the potential energy surface (PES), contains all the information about the chemical reactions that are thermally possible in that system. Searching the PES will provide predictions of all those reactions. However, a typical PES has a high dimensionality, making it too large to search exhaustively. Thus, in practice it is impossible to find all possible chemical reactions. Nonetheless, by restricting the search to a contour of constant energy (isopotential), it is possible to find at least some reactions. Thus, this technique is a useful tool to supplement the predictions of a human expert.

Results and Future Plans: Several algorithms have been designed to implement the general task of isopotential searching, including one that is well suited for large-scale parallelization. Prototype software has been written and used to test the technique. The first tests were for chemical reactions that were predicted incorrectly by experts, with the correct results discovered later. In all cases, the correct reaction mechanisms were successfully produced by isopotential searching methods. A complex example is shown in the figure. In the future, we will (1) test the procedures on other reactions that are already known but that represent different types of chemistry (e.g., transition metals), (2) apply the methods to make new predictions for important systems, and (3) distribute the software as appropriate.

Publication: