Numerical and Machine Learning Assisted Simulations of Complex Polymer Systems

Abstract

This dissertation presents the theoretical, numerical, and experimental results of studies on a varieties of complex polymer systems using a combination of analytical models, numerical simulations, and machine learning methods.

We begin with the benchmark of a new hybrid encapsulation membrane that combines an inert fluorinated polymer matrix infused with a dense omniphobic lubricating oil to protect materials with high water sensitivity while simultaneously offering mechanical flexibility, transparency and scalability. We evaluate this new protecting strategy on halide perovskite, a promising material for next-generation solar cells and photoelectronic devices. We conduct an extensive investigation by combining numerical and analytical tools to demonstrate that the lubricant infusion leads to a defect-free membranes where the water transport is dominated by the diffusion of small water cluster through a highly hydrophobic membrane.

We extend the use of numerical simulations to temperature-responsive liquid crystal molecules. In particular, we analyze end-on molecules, a particular class of liquid crystals with the ability to form covalent bonds at the extremity of the molecules. Once polymerized, these molecules are experimentally used as a cornerstone to build a new soft robotics platform for bio-inspired muscles and self-powered actuators. We demonstrate that end-on phase behaviors or dominated by thermo-activated alkyl chains conformations that dictates the molecular order. We then use molecular dynamics models to simulate how liquid crystal moieties order themselves when they are polymerized to form liquid crystal elastomers. We demonstrate that the intermolecular pi-pi stacking is the driving force behind liquid crystal mesophases. With this theoretical work, experimentalists are capable of designing temperature-responsive non-monotonic mechanical behaviors with a full mechanistical understanding of the underlying material by controlling the molecular order of end-on LCEs using external magnetic fields. To further bridge the gap between experiments and numerical simulations, we extend an existing diffraction theory to simulate X-ray scattering patterns from molecular dynamics simulations, with a high potential impact for scientists to enhance and test their understanding of materials at the molecular level.

Finally, we combine classical computation with modern machine learning tools to evaluate the recent advancement in supervised learning for molecular simulations of liquid crystals. We established an entire pipeline starting from ab-initio molecular simulations and trained our own interatomic ML potentials for bulk liquid crystal simulations. Because the resulting ML-powered MD simulations still maintain a significant computation cost, we then develop an alternative, ultrafast prediction tool using unsupervised learning and polymer theory. We show that neural networks can predict the phase transition of lyotropic liquid crystals by comparing directly with an analytical model. We hope this work will give rise to future research involving advanced predictions of complex molecular systems with ultralow latency.