Carbon Nanomaterials: Machine learning unveils carbon growth mechanism

Researchers have long sought to model how carbon atoms interact with metal surfaces to form graphene and other carbon materials. Developing this capability would elucidate the growth mechanisms of carbon nanostructures on catalytic substrates, offering new insights into controlling growth, designing novel catalysts and materials, improving energy efficiency, and enhancing scalability.

However, conventional computational methods, such as density functional theory (DFT) and kinetic Monte Carlo (KMC) simulations, have struggled to capture the full complexity of these dynamic processes.

While DFT calculations restrict simulations to small systems due to computational expense, KMC models cannot fully capture the atomic-level dynamics on evolving catalytic surfaces—hindering the development of accurate and predictive models for carbon growth.

In a 2024 article, Di Zhang, Hao Li and co-workers at AIMR developed an active machine-learning model to address these challenges¹. The team combined molecular dynamics (MD) and time-stamped force-biased (tf) MC simulations with the gaussian approximation potential to create an efficient approach for simulating carbon deposition on metal surfaces; they also integrated a smooth overlap of atomic positions-based data selection strategy to optimize the training set and improve predictive accuracy.

“A key innovation of this work is its on-the-fly machine-learning training method,” explains Zhang, first author of the article. “The model continuously refines itself by selecting the most relevant atomic configurations. This approach not only enhances predictive accuracy and maintains computational efficiency but also enables the real-time exploration of a wide range of reaction pathways.”

Replicating key processes—including carbon diffusion, chain formation, and graphene nucleation—the study revealed that carbon growth on Cu(111) surfaces follows a specific sequence, with adsorbed copper atoms playing a critical role in stabilizing carbon structures. The simulations also demonstrated that oxygen contamination can significantly disrupt graphene nucleation, aligning with experimental observations.

By leveraging machine learning, this work establishes a robust theoretical framework for designing metal catalysts to optimize carbon nanostructure synthesis. These findings lay the groundwork for more efficient, large-scale production of graphene and other carbon-based materials, accelerating their adoption in electronics and energy storage.

Personal insights from Dr. Di Zhang

Which aspect of your study’s results was the most unexpected, and why did it surprise you?

The most unexpected aspect of my study’s results was how efficiently active learning improved the training set for the machine learning force field. Traditional methods rely on random sampling, which often requires large datasets for high accuracy. By combining the MD/tfMC enhanced sampling algorithm with active learning, I streamlined the process, allowing the model to achieve high-precision predictions with significantly less training data. It was surprising to see how strategic sampling not only reduced data requirements but also maintained—or even improved—accuracy, marking a major advancement over conventional approaches.