A Neural-Network Controlled Dynamic Evolutionary Scheme for Global Molecular Geometry Optimization

Studying the molecular properties and reactivity of molecular systems requires, in a majority of cases, finding the geometric structure of a molecule corresponding to the (global) energy minimum. The issue is especially difficult in studies on nanoand biosystems. The difficulty arises from the fact that the number of local minima on the potential energy hypersurface is growing exponentially with the system size (Unger and Moult, 1993; Hendrickson, 1995; Wales, 1999). Thus, as one could suspect, the search for the optimum geometry of a molecular system belongs to the class of NP-hard problems (Unger and Moult, 1993; Hendrickson, 1995). In recent years, several attempts have been made to address the issue (Floudas and Pardalos, 2000; Pintér, 2006; Sierka et al., 2007). Still, the problem of geometry optimization cannot be considered to be satisfactorily solved, and the computational chemistry community is looking forward to a robust, reliable method for finding the global minimum of molecular potential energy. In this work we propose a new evolutionary scheme for molecular geometry optimization. The main feature of the algorithm is that a neural network is used to dynamically tune parameters of the evolutionary process. Additionally, the approach efficiently exploits domain specific features of the optimization problem. The paper is structured as follows. In Section 2 the existing approaches to global geometry optimization are concisely presented and their drawbacks are briefly discussed. A new algorithm, addressing the shortcomings of the preexisting ones, is introduced in Section 3, with the implementation details provided in Section 4. Results of calculations for selected molecular systems are shown and shortly discussed in Section 5. Conclusions are briefly drawn in Section 6.