Electron density mapping of boron clusters via convolutional neural networks to augment structure prediction algorithms.

Determination and prediction of atomic cluster structures is an important endeavor in the field of nanoclusters and thereby in materials research. To a large extent the fundamental properties of a nanocluster are mainly governed by its molecular structure. Traditionally, structure elucidation is achieved using quantum mechanics (QM) based calculations that are usually tedious and time consuming for large nanoclusters. Various structural prediction algorithms have been reported in the literature (CALYPSO, USPEX). Although they tend to accelerate the structure exploration, they still require the aid of QM based calculations for structure evaluation. This makes the structure prediction process quite a computationally expensive affair. In this paper, we report on the creation of a convolutional neural network model, which can give relatively accurate energies for the ground state of nanoclusters from the promolecule density on the fly and could thereby be utilized for aiding structure prediction algorithms. We tested our model on dataset consisting of pure boron nanoclusters of varying sizes.