Machine learning methods for aerosol synthesis of single-walled carbon nanotubes. (15th January 2023)

• Record Type: Journal Article
• Authors: Krasnikov, Dmitry V.; Khabushev, Eldar M.; Gaev, Andrey; Bogdanova, Alisa R.; Iakovlev, Vsevolod Ya.; Lantsberg, Anna; Kallio, Tanja; Nasibulin, Albert G.
• Is Part Of: Carbon. Volume 202(2023)Part 1
• Page Start: 76
• Page End: 82
• Publication Date: 2023-01-15
• Subjects: Floating catalyst CVD -- Single-walled carbon nanotube -- Machine learning -- Transparent conductive films; Carbon -- Periodicals; Carbone -- Périodiques; Koolstof; Toepassingen; Electronic journals; 546.681
• DOI: 10.1016/j.carbon.2022.10.044
• Languages: English
• ISSNs: 0008-6223

Abstract

Abstract: This work is devoted to the strategy towards the optimal development of multiparametric process of single-walled carbon nanotube (SWCNT) synthesis. Here, we examine the implementation of machine learning techniques and discuss features of the optimal dataset size and density for aerosol chemical vapor deposition method with a complex carbon source. We employ the dataset of 369 points, comprising synthesis parameters (catalyst amount, temperature, feed of carbon sources) and corresponding carbon nanotube characteristics (yield, quality, structure, optoelectrical figure of merit). Assessing the performance of six machine learning methods on the dataset, we demonstrate Artificial Neural Network to be the most suitable approach to predict the outcome of synthesis processes. We show that even a dataset of 250 points with the inhomogeneous distribution of input parameters is enough to reach an acceptable performance of the Artificial Neural Network, wherein the error is most likely to arise from experimental inaccuracy and hidden uncontrolled variables. We believe our work will contribute to the selection of an appropriate regression algorithm for the controlled carbon nanotube synthesis and further development of an autonomous synthesis system for an "on-demand" SWCNT production. Graphical abstract: Prediction error for all the methods used (as well as the experimental error) with respect to four output features of carbon. Image 1 Highlights: Machine learning applied to semi-industrial aerosol reactor for nanotube synthesis. Six methods applied to dataset of 369 points with 3 input and 4 output features. Neural networks outperform other tested algorithms including decision-tree methods. Prediction errors comparable to experimental inaccuracy achieved for all features. The dataset of as low as 250 points found to be sufficient for moderate performance.