Session: 17-15-01: Society-Wide Micro/Nano Poster Forum

Paper Number: 99101

99101 - Exploring High Thermal Conductivity Polymer Blends via Data-Driven Approach 

    Polymers are ubiquitous and massively produced nowadays because they are ideal for daily usage as well as many high-tech applications, like capacitors, Li-ion batteries, and membranes. With the massive expansion of machine learning and data-driven techniques, polymer informatics is one of the domains that has drawn attention in efficiently developing and designing desired materials. Over the course of the past few decades, various polymer datasets have been built and explored and numerous surrogate models have been trained on these datasets to predict the properties, achieving great success in identifying promising candidates in terms of required target properties. Polymer blend is a subtype of polymer that combines two or more polymers, and it has been long considered as a cost-effective and simple solution to improve or customize material properties and functionality, compared with designing a novel polymer material from scratch. However, so far, much of the focus on polymer informatics have been restricted to homopolymers, leaving the chemical space of polymer blends largely unexplored. In this work, a high-throughput molecular dynamics (MD) simulation pipeline is combined with machine learning algorithms to discover polymer blends with a higher thermal conductivity than the individual component polymers. Based on the e­xisting P­olyInfo database, the thermal conductivity of ~600 randomly selected amorphous homopolymers is first calculated using MD simulations. ~200 amorphous polymer blends with varying blending ratios (1:5, 1:1, and 5:1) are then randomly selected from this homopolymer dataset to calculate their thermal conductivity. The two datasets are then combined and employed to train a machine learning regression model to find the best representation method for polymer blends. To our knowledge, this is for the first time such an effort is done, which is based on the modification of the well-established descriptors for homopolymers. Moreover, a classification model is trained on the combined dataset to further screen a much larger polymer blends database, ~550K, generated from all the possible combinations of 2 different polymers in the homopolymer dataset with 3 different blending ratios, providing important guidance on the discovery of polymer blends with MD-computed thermal conductivity higher than the individual component homopolymers. Besides, the imbalanced nature of the combined dataset, in which only a small minority of candidates have higher thermal conductivity value than the component polymers, is handled using over-sampling techniques. In conclusion, the results and strategies from this work can be useful for extending the chemical space of current polymer informatics research to polymer blends and may contribute to the automated design of high-performance polymer blends.

Presenting Author: Jiaxin Xu University of Notre Dame

Presenting Author Biography: Jiaxin Xu is a current graduate student in the Department of Aerospace and Mechanical Engineering at the University of Notre Dame. His research focuses on material informatics and molecular dynamics simulation.

Authors:

Jiaxin Xu University of Notre Dame
Hanfeng Zhang University of Notre Dame
Tengfei Luo University of Notre Dame