(7b) Applications of Machine Learning and Bayesian Statistics in Coarse-Grained Molecular Dynamics

Historically, developing accurate coarse-grained (CG) models has been a time-consuming and arduous process. This is largely due to the absence of automated computational frameworks and methodologies to facilitate the model development. I will showcase our group’s efforts to address these grand challenges through an example of temperature and chemically transferable CG model of poly(N-isopropylacrylamide) (PNIPAM), which was used to investigate bottlebrush polymers (BBPs) of different shapes. Using these CG models, alongside conventional worm-like structures, we have created and simulated a variety of PNIPAM BBPs with distinctive shapes. This was accomplished by manipulating the length and placement of side chains along the backbone. We have created a novel analysis method based on convolutional neural networks, which can identify various structural and shape-related characteristics of BBPs that are not typically captured by traditional methods. This approach provides us with unique and valuable insights into the similarities between different BBPs. I will also provide a brief overview of our current endeavors to advance the field of CG model development by utilizing a Bayesian framework to quantify and leverage model uncertainties for further enhancements. The models and methods presented here are transferable and can be used to develop CG models of soft- and hard-materials for accelerating materials design in various fields.