Evaluation of process-structure-property relationships of carbon nanotube forests using simulation and deep learning

Abstract

This work is aimed to explore process-structure-property relationships of carbon nanotube (CNT) forests. CNTs have superior mechanical, electrical and thermal properties that make them suitable for many applications. Yet, due to lack of manufacturing control, there is a huge performance gap between promising properties of individual CNTs and CNT forest properties that hinders their adoption into potential industrial applications. In this research, computational modelling, in-situ electron microscopy for CNT synthesis, and data-driven and high-throughput deep convolutional neural networks are employed to not only accelerate implementing CNTs in various applications but also to establish a framework to make validated predictive models that can be easily extended to achieve application-tailored synthesis of any materials. A time-resolved and physics-based finite-element simulation tool is modelled in MATLAB to investigate synthesis of CNT forests, specially to study the CNT-CNT interactions and generated mechanical forces and their role in ensemble structure and properties. A companion numerical model with similar construct is then employed to examine forest mechanical properties in compression. In addition, in-situ experiments are carried out inside Environmental Scanning Electron Microscope (ESEM) to nucleate and synthesize CNTs. Findings may primarily be used to expand the forest growth and self-assembly knowledge and to validate the assumptions of simulation package. Also, SEM images can be used as feed database to construct a deep learning model to grow CNTs by design. The chemical vapor deposition parameter space of CNT synthesis is so vast that it is not possible to investigate all conceivable combinations in terms of time and costs. Hence, simulated CNT forest morphology images are used to train machine learning and learning algorithms that are able to predict CNT synthesis conditions based on desired properties. Exceptionally high prediction accuracies of R2 > 0.94 is achieved for buckling load and stiffness, as well as accuracies of > 0.91 for the classification task. This high classification accuracy promotes discovering the CNT forest synthesis-structure relationships so that their promising performance can be adopted in real world applications. We foresee this work as a meaningful step towards creating an unsupervised simulation using machine learning techniques that can seek out the desired CNT forest synthesis parameters to achieve desired property sets for diverse applications.