Neural network finds three growth recipes for the same carbon nanotube array

Researchers have utilized artificial intelligence to discover multiple growth recipes that yield carbon nanotube arrays with similar height, density, and alignment. In validation tests, 26 out of 27 growth batches achieved scores above 95% for how closely the measured arrays matched their targets. The study was published in National Science Review.

Carbon nanotubes are minuscule, hollow carbon structures. When they grow collectively on a surface, they form an array resembling a miniature forest. The height, density, and alignment of this forest influence the material's behavior, making control over these factors crucial for applications in thermal management, energy storage, and flexible electronics.

Creating arrays to precise specifications is challenging due to the interconnected nature of processing conditions. Altering temperature, gas flow, or treatment time can simultaneously affect multiple structural features, and a change that improves one aspect may negatively impact another. With eight processing variables and ten settings per variable, there are 100 million possible combinations.

The researchers developed an AI-assisted inverse-design framework to navigate this complex space. They compiled an experimental database of 200 distinct recipes for water-assisted chemical vapor deposition. A neural network learned the relationships between processing conditions and array height, density, and alignment. An optimization algorithm then worked backward from a specified structural target to identify suitable recipes.

The approach was designed to find alternatives. For each of three targets, repeated computational searches generated 200 candidate solutions. The researchers selected three recipes with significantly different processing conditions for each target and tested each recipe in three independent growth batches.

Of the 27 validation batches, 26 achieved structural descriptor matching scores above 95%, indicating how closely the measured array features matched the prescribed targets. However, the arrays still varied at a smaller scale. For one target, the three methods produced closely matched overall array features, but the mean outer diameters of the nanotubes were approximately 3.8, 4.8, and 7.2 nanometers, with differences in wall structures as well.

The researchers highlighted that the findings expand the goal of inverse synthesis from identifying a single suitable recipe to discovering multiple experimentally viable options. The framework allows exploration of routes beyond the immediate vicinity of a known recipe. While the initial database is an experimental investment, the learned relationships can be reused to screen additional targets.

Alternative recipes could eventually enable researchers to work within practical processing constraints while maintaining selected structural features. Applying this approach to other catalysts or growth methods would necessitate new experimental data and model retraining.

The study included researchers from Huazhong University of Science and Technology and Beihang University. Lei Zhu, a doctoral student at Huazhong University of Science and Technology, is the first author, with Prof. Ming Xu as the corresponding author.