Page 46 - Read Online

Zhu et al. J. Mater. Inf. 2025, 5, 8 https://dx.doi.org/10.20517/jmi.2024.76 Page 3 of 20

complex interactions between atoms.

Graph neural networks (GNNs) are recognized as powerful and flexible deep learning tools that represent
[19]
complex information through patterns and their relationships . GNNs have demonstrated excellent
performance in material property prediction. For instance, to overcome the limitations associated with
manual feature construction and complex transformations in crystal material design, Xie et al. developed
the crystal graph convolutional neural networks (CGCNN) framework to learn material properties directly
from crystal structures . It demonstrated DFT-level accuracy for predicting multiple material properties,
[20]
along with interpretable chemical insights. To enhance the accuracy and efficiency of the CGCNN in
predicting material properties, Park et al. developed an improved version (iCGCNN) by incorporating
[21]
Voronoi tessellation, three-body interactions, and optimized bond representations . The improved model
achieved a 20% increase in accuracy and a 2.4-fold rise in the success rate for high-throughput material
discovery. To address challenges in predicting material properties and designing new materials in materials
science and chemistry, Reiser et al. reviewed the fundamentals of GNNs, current applications, and potential
in accelerating simulations, material screening, and inverse design .
[22]
GNNs offer a novel approach for predicting the density of intermetallic compounds. This is achieved by
encoding the crystal structure of compounds as graph-structured data, which facilitates the learning of
[23]
information atoms and their bonding interactions . The GNN method avoids the complex feature
engineering required by traditional methods by capturing directly the information of atoms and bonding
relationships with explicit physical and chemical properties [24,25] . Furthermore, the GNN-based method
extracts structural differences between crystals using input graph structure data, which consists of
topological structures made up of atoms and bonds [26,27] . This methodology uncovers underlying
relationships between structural characteristics and material properties, highlighting the potential of GNNs
to enhance the accuracy and efficiency of material property predictions. However, existing GNN models
may encounter difficulties when directly applied to intermetallic compounds with isomeric structures.
These models are not sufficient for direct application to the density prediction of intermetallic compounds
across different crystal structures.

In this study, a new intermetallics GNN (IGNN) is constructed to perform regression and classification
predictions on the density of over 7,000 binary intermetallic compounds, and to compare its performance
with traditional machine learning models. The t-distributed stochastic neighbor embedding (t-SNE)
method is employed to visualize the IGNN’s performance in capturing differences across various crystal
structures. Furthermore, ternary and quaternary intermetallic compound density datasets are constructed as
test sets to validate the effectiveness of the proposed IGNN model. This work presents a novel and efficient
approach for predicting the densities of intermetallic compounds, offering valuable insights into materials
design and advancing the development of data-driven methods in materials science.

METHODOLOGY
Data sources and preprocessing
Density data for binary intermetallic compounds are collected using the API provided by the materials
project website . Only compounds composed of metallic elements are included, while those containing
[28]
non-metallic elements or noble gases are excluded to ensure data quality and consistency. To guarantee
thermodynamic stability, only compounds with an “energy above hull” value of 0 are retained. This
selection ensures high stability and reliability for practical use, minimizing prediction errors caused by
unstable compounds. Each downloaded data entry includes the following: Material ID, chemical formula,
crystal system, density, and crystallographic information file (CIF). CIF is a standardized format used to