From Atomistic Models to Machine Learning: Predictive Design of Nanocarbons under Extreme Conditions

E. A. Huerta, Xiaoli Yan, Millicent A. Firestone, Murat Keçeli, Santanu ChaudhuriSSRN 2025 · DOI

TL;DR

This paper uses advanced molecular dynamics simulations to understand how nanodiamonds transform into other carbon structures under extreme conditions. They then leverage this data to build a machine learning model that can predict these transformations, paving the way for designing new carbon nanomaterials.

What Problem Does This Paper Address?

The formation of valuable nanocarbon structures, like nanodiamonds, often occurs under extreme conditions such as high-explosive detonations. While we know these conditions can create nanodiamonds, the subsequent transformations during cooling and pressure release are poorly understood. This lack of understanding hinders our ability to control the synthesis of specific nanocarbon materials. This paper aims to bridge this gap by investigating how different initial nanodiamond shapes and post-detonation processing conditions (cooling and pressure release rates) influence the resulting nanocarbon structures.

Methods

The core of this research lies in GPU-accelerated reactive molecular dynamics (ReaxFF) simulations. ReaxFF is a powerful force field that allows for the dynamic formation and breaking of chemical bonds, making it ideal for studying chemical reactions under extreme conditions. The simulations model nanodiamonds (with various initial morphologies like cuboctahedral, octahedral, and hexagonal prisms) subjected to a two-stage cooling and pressure-release process, mimicking post-detonation expansion.

To make sense of the vast amount of simulation data, the researchers employed machine learning (ML) regressors, specifically a multilayer perceptron (MLP), along with other models like B-spline, gradient boosting, and random forest. These models were trained on over 10⁵ node-hours of simulation data to predict the number of graphitized layers based on temperature and pressure trajectories.

Polyhedral Template Matching (PTM) was used to classify the local atomic environments of carbon atoms into cubic diamond (CUB), hexagonal diamond (HEX), or sp² hybridized states (SP2). Ring analysis was performed to count N-membered rings (N=4-8) within the sp² carbon structures. Simulated X-ray diffraction patterns were also generated to characterize the resulting structures.

Datasets

The primary dataset used in this study is generated from over 10⁵ node-hours of ReaxFF molecular dynamics simulations. These simulations cover a range of nanodiamond morphologies and various nonlinear cooling and pressure-release trajectories. This extensive dataset was curated to train and test the machine learning models.

Key Insights and Ideas

  • Morphology Matters: The initial shape of the nanodiamond significantly influences how it transforms. For instance, octahedral nanodiamonds tend to form carbon nano-onions, while hexagonal prisms lead to parallel-stacked graphite layers.
  • Cooling vs. Pressure Release: The interplay between cooling and pressure release rates is crucial. Rapid cooling with slow pressure release generally preserves the diamond structure, while slow cooling with rapid pressure release promotes graphitization.
  • Lonsdaleite as an Intermediate: Hexagonal diamond (lonsdaleite) emerges as a transient interfacial phase during the graphitization of cubic diamond, suggesting a reverse phase transition pathway.
  • ML for Predictive Design: Machine learning can effectively predict nanocarbon transformations, enabling a data-driven approach to designing materials with desired properties. This significantly speeds up the exploration of synthesis parameter space compared to traditional simulations.

Key Results

  • The MLP model achieved a high predictive accuracy for the number of graphitized layers, with an R² of 0.893 on the training set and 0.904 on the test set.
  • Different nanodiamond morphologies lead to distinct products: cuboctahedral and octahedral nanodiamonds can form bucky diamonds and carbon nano-onions, while hexagonal prisms yield structures resembling carbon dots.
  • Rapid cooling and slow pressure release were found to be optimal for retaining the cubic diamond structure, with only a single surface layer of graphitic carbon forming.
  • Slower cooling rates, especially when combined with faster pressure release, led to more extensive graphitization and structural disorder.
  • The ML models can predict graphitization outcomes in fractions of a second, a significant speedup compared to the weeks required for MD simulations.

Limitations and Future Work

  • Timescale Disparity: The MD simulations operate on picosecond timescales, which are much faster than the microsecond to millisecond timescales of experimental detonations. This means the simulations capture relative trends rather than absolute experimental kinetics.
  • Simplified Conditions: The simulations neglect bulk and surface defects, variations in surface chemistry, and chemical impurities, which are present in real detonation environments and can influence transformation thresholds.
  • Particle Size: The study primarily focuses on nanodiamond particles in the 2-4 nm range, emphasizing surface and curvature effects. Larger particles might behave differently.
  • Future Work: The authors suggest incorporating initial nanodiamond crystal structure (morphology) into the ML models for more comprehensive predictions. They also propose scaling the framework to chemically heterogeneous systems and including coupled shock-release dynamics to better represent detonation environments. Investigating the role of explicit shear loading is also highlighted as a critical avenue for future research.

Summary updated

BibTeX
@misc{huerta2025from,
  author = {E. A. Huerta and Xiaoli Yan and Millicent A. Firestone and Murat Keçeli and Santanu Chaudhuri},
  title = {{From Atomistic Models to Machine Learning: Predictive Design of Nanocarbons under Extreme Conditions}},
  year = {2025},
  doi = {10.2139/ssrn.5776082},
  howpublished = {SSRN},
}
Figure