Relationship of structure and mechanical property of silica with enhanced sampling and machine learning

YP Deng and T Du and H Li, JOURNAL OF THE AMERICAN CERAMIC SOCIETY, 104, 3910-3920 (2021).

DOI: 10.1111/jace.17779

As one of the most abundant materials on Earth, silica has been widely studied in various crystal structures and glassy states. However, in terms of molecular structure and the corresponding mechanical properties between those of ordered and disordered states, partially ordered silica states are not well-explored in the relevant literature owing to a low probability of appearance in experiments and simulations. The lack of this knowledge significantly hinders the understanding of the inherent mechanism of mechanical properties and limits the applications in many engineering fields. In this study, we present an exploration of the complex interdependent relations of the structural properties of silica over a wide range of free energy surfaces and establish a machine learning-based prediction model via high-throughput molecular dynamics simulations coupling with an enhanced sampling method. Each scale of structure information of samples with varying crystallinity is analyzed. First, descriptors of silica structures were identified and selected as inputs to a deep neural network (DNN). The results indicate that our DNN-based approach can provide an accurate prediction of bulk modulus, shear modulus, and tensile strength of silica samples. Furthermore, the generalizability of the machine learning model is verified on the prediction tasks for much larger silica systems, as well as silica quenched at varying cooling rates. Overall, the enhanced sampling method can reliably accelerate the exploration of free energy surfaces and collection of training samples, and machine learning methods are effective in generating accurate and reliable predictions of mechanical properties of materials over the free energy surface.

Return to Publications page