Hybrid quantum-classical approach for atomistic simulation of metallic systems

The learn-on-the-fly (LOTF) method [G. Csanyi et al., Phys. Rev. Lett. 93, 175503 (2004)] serves to seamlessly embed quantum-mechanical computations within a molecular-dynamics framework by continual local retuning of the potential's parameters so that it reproduces the quantum-mechanical forces. In its current formulation, it is suitable for systems where the interaction is short-ranged, such as covalently bonded semiconductors. We propose a substantial extension of the LOTF scheme to metallic systems, where the interaction range is longer and the many-body nature of the potential prevents a straightforward application of the original LOTF technique. We propose to realize the force optimization stage in a divide-and-conquer fashion and give detailed analysis of the difficulties encountered and the means to overcome them. We show how the technique, which we have termed divide and conquer learn-on-the-fly, can be parallelized to utilize several tens of processors. Finally, we present the results of an application of the proposed scheme (utilizing tight binding for the quantum-mechanical part) to nanoindentation and nanoscratching of single-crystal Cu.