Design of mechanically robust nanoporous silicon nitride membranes via combined molecular dynamics and generative deep learning approach

Ali K. Shargh, PhD Qualifying Exam, Advised by Niaz Abdolrahim

Friday, April 9, 2021
1:30 p.m.

Nanoporous silicon nitride membranes (NPN) are freestanding ultrathin films that are finding diverse applications in biological separations, DNA translocation, and hemodialysis thanks to their nanoscale thickness as well as high porosity. The most common version of NPN, available commercially, is produced by using rapidly crystallized amorphous silicon films as templates which is first introduced in McGrath Laboratory at University of Rochester. It has been concluded from experiments that the range of working pressure and the size of those NPN in applications are controlled by their mechanical properties. Thus, understanding how the microstructure of NPN controls their mechanical properties will be the key to design of robust, large-area membranes for improved future applications. This research work will develop such fundamental understandings using a wide range of computational techniques such as: molecular dynamics (MD), finite element (FE) as well as deep learning modeling in close collaboration with experiments.  


The initial aim is focused on atomistic modeling of the NPN in crystalline and amorphous phases. Effects of primary microstructural parameters including porosity, pore size and pore distribution on mechanical properties as well as deformation mechanisms of NPN is explored. Here, it is concluded that the pore distribution is the dominant parameter that controls plastic deformation mechanisms of NPN. More specifically, different types of shear band networks are captured based on the pore distribution, and it is shown that the type of shear band network ultimately controls the failure behavior and thus the ductility of NPN. In addition, it is unraveled that while those shear bands reside in the slip system of crystalline NPN, the shear bands propagate along specific directions of amorphous NPN in agreement with Mohr-Coulomb yield surface theory. A mathematical model based on Griffith’s theory is developed thereafter to predict the strength of NPN based on their pore distribution as well as porosity. Some of the central observations of MD simulations is further confirmed from bulge test experiments. For instance, it is concluded from experiments that the strength of NPN decreases with the increase of porosity in accordance with MD simulations. In the second aim, more complex microstructural parameters than the conventional ones such as stochastic clustering of the pores will be included into the investigation which urges to employ deep neural networks integrated with FE simulations to extend the fundamental understandings on the correlation of mechanical properties of NPN with their microstructures. Moreover, the dominant microstructural parameters in controlling the mechanical properties of NPN among all the candidates will be highlighted from statistical analysis. In the final aim, microstructures of NPN with optimized mechanical properties will be designed employing an optimization scheme consisting of convolutional neural networks.