Artificial Intelligence-Driven Prediction for Efficient CNT Growth Simulation

Carbon nanotube (CNT) forests are comprised of vertically oriented CNTs frequently synthesized to hundreds of microns of length. Because of morphological imperfections and continuous CNT delamination during synthesis, the ensemble properties of CNT forests are often orders of magnitude less than what is predicted from the properties of individual CNTs. Concurrently, the parametric space for CNT forest synthesis is inexhaustible. In a conventional chemical vapor deposition synthesis, the parameter space includes combinations of catalyst film composition, catalyst buffer layer composition, growth temperature, carbon feedstock, water vapor concentration, pressure, carrier gas, among others. The conventional synthesis and characterization process is expensive and prone to error. Artificial Intelligence (AI) driven experimental exploration, as demonstrated by the AFRL Autonomous REsearch System (ARES), has demonstrated a drastic acceleration in determining fundamental single-walled CNT synthesis mechanisms. In this work, we explore machine learning predictions versus traditional Finite-Difference Time-Domain (FDTD) based simulations to expedite the numerical exploration process of CNT forests.<br/>Simulating the growth and assembly of dense forest CNTs has been demonstrated using a time resolved finite element simulation. The simulation considers distributed CNT growth rates, diameters, the interactions between contacting CNTs, and mechanical delamination of CNTs from the growth substrate. After CNT forest growth and assembly, the simulation may be used to simulate the mechanical, electrical, and thermal properties of the forest. The simulation is viewed as increasingly important for applications such as CNT interface materials, composite materials, and electron emitters. Because of the low cost and relative speed of simulations, the numerical approach may be a means to accelerate the search of process-structure-property relationships for custom applications. However, conventional methods become increasingly computationally expensive as CNT forest structures increase in length and complexity and may limit the application of numerical approaches.<br/>To address this limitation, we present a custom recurrent neural network (RNN) pipeline for efficient CNT growth prediction. By leveraging time-varying data from past simulations, the proposed model incorporates growth positional information directly. In the computational synthesis of CNTs, two principal forces are critical - internal elastic forces within each CNT influence the direction and height of growth, and the external van der Waals force determines the attractions and repulsions between different CNTs. These forces inherently lead to the collective deformation of growing CNTs. Unlike traditional RNNs that consider static time evolving signals, the proposed method not only forecasts the growth of the CNTs but also considers changes in the forest structure due to internal CNT-CNT interactions. Compared to iterative physics-based simulations, this framework reduces computational costs, facilitating more efficient search of CNT growth parameters. While applied to CNT growth, the proposed method shows promise for a diverse range of growth phenomenon where underlying structures change during the growth process. This methodology demonstrates replacing computationally expensive iterative simulation with machine learning emulation.