Dr. Bogdan Palosz

Abstract:

The atomic structure of nanocrystals differs from that of bulk crystals both on the surface and inside the grains.  To investigate the unique properties of nanomaterials, innovative tools specifically designed for structural studies of nanomaterials are needed. 

Complete information on the size, shape, and atomic structure of nanocrystals is contained in diffraction data. Since the atomic structure of an individual grain depends on its size and shape, conventional analytical tools of crystallography designed for polycrystalline materials are clearly insufficient. Therefore, creation of a subfield of crystallography, nanocrystallography, is currently under consideration that being recommended specifically for structural studies of nanomaterials [1,2].  An advanced software and methodology designed for elaboration of diffraction data of nanocrystalline materials were proposed that can be used to identify their actual structure [3,4]. However, this methodology is labor-intensive and time-consuming, is used only sporadically, and is unlikely to become the standard method recommended for characterizing the structure of nanomaterials. It is therefore worth considering the implementation of artificial intelligence (AI) algorithms in nanocrystallography. AI is used for fast and reliable classification of information contained in large data sets, what in our case means search through a large number of tentative structural models of nanograins [5]. For our studies machine learning (ML) has been chosen to facilitate the processing of the diffraction data. ML is able to effectively analyze the network of connections between different parts and features of given object which for us is the diffraction data.  Diffraction patterns of grains of similar size and different shapes show similar shapes and positions of Bragg peaks and reveal only litle differences that are difficult to quantify. The key question is whether ML classifiers are able to distinguish between grains by comparing their representations in reciprocal and real spaces, S(Q) and G(r). The advantage of ML over other numerical techniques for statistical analysis stems from its ability to discover relationships between objects completely independently during the training stage. In our case, the actual objects under study are individual grains and their assemblies, and the objects used for ML study are their structural factor S(Q) and interatomic distance functions G(r).

The amount of data available for processing and training is crucial for the performance of the ML algorithm. Since the availability of experimental diffraction data is very limited, we use theoretical data calculated for nanocrystal models ranging in size from 2 to 15 nm, with different shapes and surfaces. DFT and MD simulations were used to relax initially perfect crystal lattice, generate surface and internal strains occurring in grains and create reliable atomic models of nanocrystals. The database consists of several thousand S(Q) and G(r) functions that are used for training. The optimal machine learning algorithms used are based on supervised techniques such as random forests (RF) and neural networks (NN). 

In real material, there is a variety of grains of different sizes and shapes.  The relevant question for our study is what are dimensions and shapes of the grains in a given material and what types of surfaces are most common. Examples of the application of the AI algorithm to analyze the appearance of grains of a specific shape and surface will be presented for real of nanocrystalline diamond samples.