Artificial neural network based transferable interatomic potentials : application for gold nanoparticles and its alloys

Abstract

Machine learning has the ability to solve a problem which is beyond the processing of a human brain. The learning and the processing of data by a machine provides different connections present within a database. Artificial neural network (ANN) is one of the robust machine learning technique that mimics the learning process of a human brain using the basics of a perceptron model. In the last few decades, ANN has been used to solve complex problems in chemistry. The calculation of potential energy surface (PES) of a nanoparticle (Np) is one such problem which can be solved using ANN. The ab initio methods can be applied for accurate prediction of PES, but their high computational cost makes them a poor alternative. ANN provides a solution to bridge the gap between accuracy of ab initio methods and low computational costs. ANN interatomic potentials provides a cheap and accurate alternative to study the structural dynamics of metallic Nps. Metallic Nps have a variety of applications which makes them an important topic to study theoretically. The study of structural dynamics of Nps leads to major insights into vacancy defects, surface energy, mechanical properties, plasmon-resonance behavior. In this thesis, the ANN interatomic potentials is constructed for gold Nps and their alloys. As gold shows a rugged PES due to relativistic effects, the fitting of PES was possible with high dimensional ANN. For constructing a PES using ANN, one of the important part is the descriptors of the atomic environment. Higher order invariants- Power Spectrum and Bispectrum have been applied with modified atomic environment density for describing the atomic environments. A transferable approach for fitting PES of an alloy system was not done prior to the work done in this thesis. For an alloy system, the PES is fitted using a single ANN by following a strategy of decoupled fitting of energy and forces. The elements are differentiated between each other using weightings in the descriptors. The PES fitted for small and medium sized clusters(∼1.8 nm) is found to be transferable to larger size clusters(<3.3 nm). The computational time for accurate calculation of energy and forces using power spectrum-ANN for a Au147 cluster was reduced to seconds, when compared to DFT (∼ hours) (calculation done on parallelized 8 CPU [GenuineIntel 2600.0 MHz]). Due to an accurate prediction of PES of gold Nps, a symmetric core evolution with increase in size of gold Nps is studied. It is observed that an icosahedron core is evolving from Au160 to Au327 to Au571. It is also observed that magic number clusters- Au147, Au309, Au561 and Au923 prefer amorphous structure over symmetric structures. The unusual bonding in gold leads to modification of the structural preference in magic number clusters. Overall, the proposed descriptors and various new models have proved to be of great significance in fitting the PES of complex system like gold.