Synthetic trade-offs exist in the synthesis of single-walled carbon nanotube (SWCNT) forests, as growing certain desired properties can often come at the expense of other desirable characteristics such as the case of crystallinity and growth efficiency. Simultaneously achieving mutually exclusive properties in the growth of SWCNT forests is a significant accomplishment, as it requires overcoming these trade-offs and balancing competing mechanisms. To address this, we trained a machine-learning regression model with a set of 585 "real" experimental synthesis data, which were taken using an automatic synthesis reactor. Subsequently, 16000 exploratory "virtual" experiments were performed by our trained model to examine potential routes toward addressing the current crystallinity-height trade-off limitation, and suggestions on growth conditions were predicted. Importantly, additional validation using "real" experimental syntheses showed good agreement with the predictions as well as a 48% increase in growth efficiency while maintaining the high crystallinity (G/D-ratio). This highlighted the effectiveness and accuracy of the predictive capability of our machine-learning model, which achieved improved results in less than 50 validation tests. Furthermore, the trained model revealed the surprising importance of the nature of the carbon feedstock, particularly the reactivity and concentration, as a route for overcoming the trade-off between the SWCNT crystallinity and growth efficiency. These results of the high-efficiency synthesis of highly crystalline SWCNT forests represent a significant advance in overcoming synthetic trade-off barriers for complex multivariable systems.
Keywords: carbon nanotube; crystallinity; growth efficiency; machine learning; trade-off.