Creating semi-Quanta multi-layer synthetic CNT images using CycleGAN

We address the challenge of creating semi-Quanta images of vertically-oriented carbon nanotubes (CNT). Our previous method, CNTNeXt, successfully predicted the mechanical properties for semi-Scanning Electron Microscopy (SEM) images of CNT forests. However, for Quanta images, a more sophisticated approach is required. We propose a novel pipeline that modifies multi-layer synthetic (MLS) CNT images, which are synthetic representations of CNT structures, to resemble Quanta images. This is achieved using CycleGAN, a deep learning model designed for image-to-image translation tasks, which can learn mappings between two different domains without paired training data. By training the CycleGAN on both Quanta and MLS images, we are able to transfer the texture of Quanta images to the MLS images. The resulting ’semi-Quanta’ images are ready to train CNTNeXt, which effectively predicts the mechanical properties, including buckling load and stiffness of synthetic MLS images. Our method represents an advancement towards our ultimate goal of creating an automated CNT growth process, and opens up possibilities for further improvements and applications.

I. Introduction

The journey from the initial discovery of new materials to their commercialization is typically a protracted and labor-intensive endeavor, often spanning a decade or two of rigorous research and development efforts before a material reaches the marketplace [1]. Traditional methods depend heavily on empirical approaches and the experiential insights of researchers to navigate the complex landscape of material synthesis, which is characterized by a multitude of variables. Deep learning has revolutionized the field of artificial intelligence, with its applications permeating a wide array of domains of image and video recognition [2], data classification [3], [4], and more. The incorporation of machine learning (ML) and artificial intelligence (AI) into this process has the potential to significantly expedite material innovation, diminishing the reliance on human heuristics and potentially revolutionizing the approach to material science research. Carbon nanotubes (CNTs) serve as a prime example of a material that has undergone this extensive development cycle, highlighting the transformative impact that ML and AI could have on the field [5], [6]. The versatility and exceptional properties of CNTs have captured the interest of numerous researchers, leading to a substantial body of work dedicated to understanding and harnessing their potential. For instance, Najmi et al. have explored effects of CNTs on termal behaviour of honeycomb sandwich panels and polymer composites [7], [8]. Haddadi investigated the friction phenomenon in carbon nanotubes [9]. Nguyen et al. have explored deep learning models for segmentation and layer decomposition of carbon nanotubes in SEM imagery [10], [11]. Ferdousi studied the mechanical characteristics of single-walled CNTs. [12], [13].