Session: 03-06: Materials Development Using Artificial Intelligence

Paper Number: 137577

137577 - Application of Machine Learning to the Design of Carbon Nanotube Bundle Microstructures via Genetic Algorithms and Convolutional Neural Networks 

In recent years, machine learning (ML) has been widely applied as a potential tool for improving understanding of material behavior. With the ability to efficiently process large amounts of data and make rapid predictions based on that data, ML has the potential to assist in the understanding of a complex material system: carbon nanotube (CNT) composites. Here, ML is applied to the design of CNT bundle microstructures, one of many multiscale building blocks of CNT composites.

The multiscale nature of CNT composites is a major challenge in the understanding and design of this material system. At the nanoscale, carbon nanotubes (CNTs) have been demonstrated to feature many desirable properties, including high tensile stiffness and strength. At the mesoscale, CNT yarns are composed of bundles of individual CNTs, or CNT bundle microstructures. The mechanical behavior of a CNT bundle microstructure is influenced by features from multiple length scales, including CNT shape at the nanoscale, bundle shape from the mesoscale, and interactions between bundles at the mesoscale. The number and inherent complexity of design features that influence CNT bundle microstructures leads to a large design space that must be searched to understand the material behavior. 

Machine learning is a potential tool to assist in the exploration of the large design space of CNT bundle microstructures. Building on the physics-based data obtained from finite element (FE) simulations, machine learning models can be trained to rapidly predict behavior of unsimulated CNT bundle microstructures. As compared to FE simulations alone, using a hybrid of FE simulations and ML predictions can increase the efficiency of exploring the large design space.

In this work, a convolutional neural network (CNN) is combined with a genetic algorithm (GA) to efficiently search the CNT bundle microstructure design space. An FE simulation framework is used to generate training data for a CNN. The CNN is trained to predict effective elastic properties of CNT bundle microstructures. The trained CNN is incorporated into a GA to form a CNN-informed GA, in which the CNN predictions are used to determine the fitness of bundle microstructures selected by the GA. The CNN-informed GA is used to design CNT bundle microstructures with desired bulk elastic responses.

The performance of the CNN-informed GA in the design of CNT bundle microstructures is assessed through a series of studies. Inspection of simulation data shows that FE simulations capture key physics-based trends in bulk microstructure behavior. Comparison of actual and predicted bulk elastic properties shows that the trained CNN can predict bulk elastic moduli with reasonable accuracy. Follow-up FE simulations of GA-identified microstructures show that the CNN-informed GA can design bundle microstructures that achieve specified objectives with reasonable accuracy. Comparison of CNN-informed GA and brute force search results shows that the CNN-informed GA finds solutions that outperform most of the brute force search results in a fraction of the time. Overall, the CNN-informed GA is found to be acceptably effective at designing CNT bundle microstructures using ML.

Presenting Author: Ibrahim Guven Virginia Commonwealth University

Presenting Author Biography: Ibrahim Guven is an Associate Professor of Mechanical and Nuclear Engineering at Virginia Commonwealth University (VCU). He was an Assistant Professor of Materials Science and Engineering at The University of Arizona. Ibrahim spent two summers as a Faculty Fellow at the Air Force Research Laboratory. He was a Visiting Professor at the University of Rennes I, France, multiple times. Ibrahim is a recipient of the NASA Group Achievement Award for outstanding work in developing materials for space exploration, which was awarded to participants of the collaborative project he worked on, US-COMP Space Technologies Research Institute.