Impressive computational acceleration by using machine learning for 2-dimensional super-lubricant materials discovery

Marco Fronzi, Mutaz Abu Ghazaleh, Olexandr Isayev, David A. Winkler, Joseph G. Shapter, Michael J. Ford

Abstract

The screening of novel materials is an important topic in the field of materials science. Although traditional computational modeling, especially first-principles approaches, is a very useful and accurate tool to predict the properties of novel materials, it still demands extensive and expensive state-of-the-art computational resources. Additionally, they can be often extremely time consuming. We describe a time and resource-efficient machine learning approach to create a large dataset of structural properties of van der Waals layered structures. In particular, we focus on the interlayer energy and the elastic constant of layered materials composed of two different 2-dimensional (2D) structures, that are important for novel solid lubricant and super-lubricant materials. We show that machine learning models can recapitulate results of computationally expansive approaches (i.e. density functional theory) with high accuracy.

Introduction

Friction and wear account for a significant portion of global energy consumption and mechanical failure. While liquid lubricants are standard in many applications, extreme environments—such as space exploration or high-temperature manufacturing—require robust solid lubricants. 2D materials like graphene and molybdenum disulfide ($MoS_2$) are premier candidates for these roles because their layers can slide past each other with minimal resistance, a phenomenon known as super-lubricity.

The primary challenge in discovering new 2D lubricants is the vastness of the "chemical space." With thousands of known monolayers, stacking them into bilayer heterostructures results in nearly 18 million possible combinations. Predicting the stability and friction properties of these combinations using traditional first-principles methods like Density Functional Theory (DFT) would require centuries of supercomputer time. This research demonstrates a machine learning (ML) pipeline designed to bypass this "computational wall."

Methodology

The researchers developed a multi-stage pipeline to move from raw crystal structures to high-fidelity property predictions:

  1. Data Generation: A small, representative subset of ~280 bilayers was calculated using DFT to provide ground-truth data for Interlayer Energy (IE) and the elastic constant ($C_{33}$).
  2. Structural Descriptors: Crystal structures were converted into mathematical graphs, initially generating over 2,700 descriptors per layer.
  3. Feature Selection: To prevent overfitting, the team used Genetic Algorithms (GA) and LASSO regression to prune the descriptors down to the most relevant 42 for energy and 89 for elasticity.
  4. Bayesian Neural Network (BNN): Unlike standard networks, the BNN provides a confidence interval for its predictions, allowing the model to estimate its own certainty.
  5. Clustering: K-means clustering ensured the training data covered a diverse range of material types, preventing bias toward common elements.
The machine learning pipeline (a, b) and UMAP visualizations of the descriptor space (c) used for screening 2D bilayer heterostructures.

Results and Discussion

The BNN model achieved impressive accuracy, with an $R^2$ of 0.80 on test sets. The average error was remarkably low: approximately 4% for interlayer energy and 11% for the elastic constant. This high performance allowed the model to successfully extrapolate properties for the entire 18-million-structure database.

Key Insights

  • Property Independence: A surprising finding was that Interlayer Energy (how "sticky" layers are) and the Elastic Constant (how "squishy" the material is) are almost entirely uncorrelated. This suggests that low friction and mechanical softness must be optimized independently.
  • Stability Filtering: By applying thermodynamic filters, the researchers narrowed the 18 million possibilities down to ~300,000 "manufacturable" bilayers, with 60,910 candidates predicted to be stable at room temperature.
  • Top Candidates: The model identified specific high-performing pairs, such as $As_4S_6-In_2Se_2$, which exhibit potential for ultra-low friction even under high-temperature conditions.
Predicted property heatmaps and relative error distributions for Interlayer Energy (IE) and elastic constant ($C_{33}$), alongside joint distributions showing the lack of correlation between the two properties.

Conclusion and Future Work

This study proves that machine learning can accurately recapitulate complex quantum mechanical calculations across a massive domain. While the model currently assumes a 0-degree twist angle between layers—ignoring the complex "moiré effects" that occur during rotation—it provides a powerful shortcut for materials discovery. Future work will focus on incorporating rotational degrees of freedom and the experimental synthesis of the identified "super-lubricant" candidates to confirm their real-world performance.

Summary updated

Abstract

The screening of novel materials is an important topic in the field of materials science. Although traditional computational modeling, especially first-principles approaches, is a very useful and accurate tool to predict the properties of novel materials, it still demands extensive and expensive state-of-the-art computational resources. Additionally, they can be often extremely time consuming. We describe a time and resource-efficient machine learning approach to create a large dataset of structural properties of van der Waals layered structures. In particular, we focus on the interlayer energy and the elastic constant of layered materials composed of two different 2-dimensional (2D) structures, that are important for novel solid lubricant and super-lubricant materials. We show that machine learning models can recapitulate results of computationally expansive approaches (i.e. density functional theory) with high accuracy.
BibTeX
@misc{fronzi2019impressive,
  author = {Marco Fronzi and Mutaz Abu Ghazaleh and Olexandr Isayev and David A. Winkler and Joseph G. Shapter and Michael J. Ford},
  title = {{Impressive computational acceleration by using machine learning for 2-dimensional super-lubricant materials discovery}},
  year = {2019},
  eprint = {1911.11559},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.11559},
}
Figure