Predicting properties of hard-coating alloys using ab-initio and machine learning methods

Henrik Levämäki, Ferenc Tasnádi, Davide G. Sangiovanni, L.j.s. Johnson, Rickard Armiento, Igor A. AbrikosovarXiv 2021 · DOI · arXiv

Abstract

Accelerated design of novel hard coating materials requires state-of-the-art computational tools, which include data-driven techniques, building databases, and training machine learning models against the databases. In this work, we present a development of a heavily automated high-throughput workflow to build a database of industrially relevant hard coating materials, such as binary and ternary nitrides. We use Vienna Ab initio Simulation package as the density functional theory calculator and the high-throughput toolkit to automate the calculation workflow. We calculate and present results, including the elastic constants, one of the key materials parameter that determines mechanical properties of the coatings, for X(1-x)Y(x)N binary and ternary nitrides, where X,Y in {Al, Ti, Zr, Hf} and fraction x = 0, 1/4, 1/2, 3/4, 1. We explore ways for ML techniques to support and complement the designed databases. We find that the crystal graph convolutional neural network model trained on Materials Project data for ordered lattices has sufficient prediction accuracy for the disordered nitrides, suggesting that the existing databases provide important data for predicting mechanical properties of qualitatively different type of material systems, in our case hard coating alloys, not included in the original dataset.

Introduction

Hard coatings, such as Titanium Nitride (TiN), are essential for protecting industrial tools from wear and heat. However, designing new "super-hard" coatings is traditionally a slow, trial-and-error process. While computational methods like Density Functional Theory (DFT) can predict material properties with high precision, they are computationally expensive, particularly for complex ternary or quaternary alloys.

A significant challenge in this field is the data gap between existing materials databases and industrial reality. Most databases, such as the Materials Project, focus on ordered compounds with fixed, repeating atomic positions. In contrast, high-performance industrial coatings are often disordered alloys, where different metal atoms are randomly substituted into the lattice. This work addresses whether machine learning (ML) models trained on ordered data can bridge this gap and accurately predict the properties of disordered systems.

Methodology

The researchers developed a heavily automated high-throughput workflow to generate a new database of industrially relevant nitrides ($X_{1-x}Y_xN$ where $X,Y \in \{Al, Ti, Zr, Hf\}$).

  • Automated DFT: The Vienna Ab initio Simulation Package (VASP) was managed by the High-Throughput Toolkit (httk) to automate thousands of calculations across three common crystal structures: Rocksalt (B1), Zincblende (B3), and Wurtzite (B4).
  • Modeling Disorder: To simulate disordered alloys within the constraints of periodic boundary conditions, the team used Special Quasirandom Structures (SQS). These small supercells are mathematically designed to mimic the local atom-to-atom correlations of a truly random alloy.
  • Graph Neural Networks: The study employed a Crystal Graph Convolutional Neural Network (CGCNN). This architecture represents a crystal as a graph where atoms are nodes and bonds are edges, allowing the network to learn structural and chemical patterns directly from the atomic arrangement.
The CGCNN architecture converts crystal structures into graphs to learn chemical-structural patterns for property prediction.

Key Results and Discussion

The study utilized a "zeroth-order" transfer learning approach: training the CGCNN on a massive dataset of ~8,000 ordered compounds from the Materials Project and testing its ability to generalize to the newly generated disordered nitride dataset without further fine-tuning.

ML Accuracy and Generalization

The most significant finding was the model's remarkable ability to generalize. Even though it had never seen disordered alloys during training, the CGCNN predicted the bulk and shear moduli of the new nitrides with high accuracy. For relaxed B3 structures, the Mean Absolute Error (MAE) was only 5.3 GPa for the bulk modulus ($B$) and 7.9 GPa for the shear modulus ($G$).

Comparison of DFT-calculated vs. ML-predicted elastic properties for both the Materials Project test set and the new disordered nitride dataset.

The Role of Atomic Relaxation

The research highlighted that atomic relaxation—allowing atoms to settle into their equilibrium positions—is critical for accurately predicting the shear modulus ($G$). Shear distortions often break local symmetry, requiring significant atomic movement to reach stability. In contrast, the bulk modulus ($B$) was less sensitive to these local adjustments.

Metastability Challenges

The study also noted that many hexagonal (B4) structures were inherently unstable, often relaxing into a planar "Bk" phase. This underscores the difficulty of designing coatings that are frequently synthesized in far-from-equilibrium states, such as those produced via physical vapor deposition.

Conclusion

This work demonstrates that ML models trained on large datasets of perfectly ordered crystals can serve as powerful tools for predicting the mechanical properties of complex, disordered alloys. While the model showed some limitations—such as underestimating the shear modulus for the B1 phase and struggling with underrepresented hexagonal phases—the overall success of this transfer learning approach is a major step forward for accelerated materials discovery. Future work will focus on fine-tuning these pre-trained models with small amounts of alloy-specific data to further enhance predictive precision.

Summary updated

Abstract

Accelerated design of novel hard coating materials requires state-of-the-art computational tools, which include data-driven techniques, building databases, and training machine learning models against the databases. In this work, we present a development of a heavily automated high-throughput workflow to build a database of industrially relevant hard coating materials, such as binary and ternary nitrides. We use Vienna Ab initio Simulation package as the density functional theory calculator and the high-throughput toolkit to automate the calculation workflow. We calculate and present results, including the elastic constants, one of the key materials parameter that determines mechanical properties of the coatings, for X(1-x)Y(x)N binary and ternary nitrides, where X,Y in {Al, Ti, Zr, Hf} and fraction x = 0, 1/4, 1/2, 3/4, 1. We explore ways for ML techniques to support and complement the designed databases. We find that the crystal graph convolutional neural network model trained on Materials Project data for ordered lattices has sufficient prediction accuracy for the disordered nitrides, suggesting that the existing databases provide important data for predicting mechanical properties of qualitatively different type of material systems, in our case hard coating alloys, not included in the original dataset.
BibTeX
@misc{levamaki2021predicting,
  author = {Henrik Levämäki and Ferenc Tasnádi and Davide G. Sangiovanni and L.J.S. Johnson and Rickard Armiento and Igor A. Abrikosov},
  title = {{Predicting properties of hard-coating alloys using ab-initio and machine learning methods}},
  year = {2021},
  eprint = {2111.01111},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2111.01111},
}
Figure