Materials design of polycarbonates at the atomistic scale with machine learning

Project began in 2021.

Polycarbonate (PC) materials are ubiquitous because of their excellent transparency, low weight, and good strength and impact resistance. Autonomous vehicles are particularly dependent upon such materials because they provide structural strength for the vehicle front end while being radio transparent, enabling sensors to assess the vehicle environment without being exposed. Light Imaging Detection and Radar (LIDAR) sensors mounted in the vehicles front and side are usually protected by polycarbonate panels which, while opaque to visible light, are transparent to the radar signals. PC also serves as a good thermal insulator which helps protect sensitive electronics in harsh environments. Finally, PC’s structural health can be assessed relatively easily by looking for micro-cracks in the structure, opening the possibility of using a system like LIDAR to automatically assess structural damage. Because PC satisfies well these many functions at once, it is one of the most common materials used in design of state-of-the-art autonomous vehicles.

Molecular dynamics (MD) modeling of polycarbonates and other polymers at the atomistic scale has proven to be a valuable tool to uncover mechanical and chemical pathways during deformation and temperature changes. However, while MD is a great tool for probing the origin of mechanical behavior, it has been so far unsuccessful as a predictive design tool due to limited quantitative accuracy and scale. The extremely limited spatial and temporal scales of MD produces known artifacts. Even when these are accounted for, the underlying physics of the atomic scale force fields cannot be fit with sufficient fidelity by empirical models to produce results which match experimental data to better than approximately a factor of 2.

Recently, new innovations in machine learning (ML) have led to several new atomistic models trained by neural networks. These potentials have the major advantage that they fit the material behavior of their training data with accuracy limited only by the dimensions of the architecture and size of the training dataset, both of which can be iterated until an acceptable tolerance is reached. A second advantage of ML atomistic models is that the generation of the potential can be nearly entirely automated, unlike the more manual process used to hand-tune traditional empirical models. Conversely, ML models require two orders of magnitude more training data than empirical models. This data is generally obtained by DFT simulations. Thus, the process of creating ML potentials requires much less human time than empirical potentials, but much more computer time. This computer time is highly parallelizable, so as computational threads become increasingly more available, it is not seen as a major impediment.

This proposed research suggests using recent advances in material modeling with machine learning (ML) to create a predictive model at the atomistic scale for the mechanical behavior of PCs. The primary question is can a neural network be trained to predict the mechanical behavior of a PC composite including strength-enhancing additives. Additionally, can the network incorporate the role of anisotropy, e.g., orientation of nanotubes, and how they affect macroscopic properties like the ease of different cracking modes and the mechanical strength under various loading conditions. This work focuses on probing additives which interact mechanically rather than chemically with the PC.

We aim to create a more complete atomic scale description of the mechanical behavior of PC composites, increasing our fundamental understanding of how to optimize impact resistance, strength, and retain necessary ductility. Moreover, it will be an improvement over existing models because our model approach enables greater reliability of results.

#3.18