A Transformer-Based GAN Architecture for Simulating Carbon Nanotube Structures in the Frequency Domain

Trevor Bohl, Minasadat Attari, Filiz Bunyak, Matthew Maschmann

Abstract: The design and discovery of new materials with specific properties is a costly and time-consuming process because of the vast search space and the high cost of experimental testing. Simulation and machine learning offer promising solutions to address these challenges. Carbon nanotubes (CNTs) are highly promising nanoscale materials renowned for their exceptional mechanical, electrical, and thermal properties, and have a wide range of potential applications. The properties of CNTs are related to their structural characteristics, which are determined by the growth parameters. In this paper, we present CNT-Former, a generative adversarial network (GAN) using a transformer-based architecture and frequency domain encoding to simulate CNT structures based on given growth parameters. Our ultimate objective is to identify the optimal growth parameters for the desired structures and properties. This approach provides a more scalable alternative to traditional finite element simulators while maintaining high accuracy in simulating CNT structures. Additionally, it allows for the integration of real, multi-modal experimental data, grounding the simulations in actual experimental results. Experimental results demonstrate promising structural fidelity and strong class discrimination capabilities for CNT-Former.