Developing an improved Crystal Graph Convolutional Neural Network framework for accelerated materials discovery

The recently proposed crystal graph convolutional neural network (CGCNN) offers a highly versatile and accurate machine learning (ML) framework by learning material properties directly from graph-like representations of crystal structures (“crystal graphs”). Here, we develop an improved variant of the CGCNN model (iCGCNN) that outperforms the original by incorporating information of the Voronoi tessellated crystal structure, explicit 3-body correlations of neighboring constituent atoms, and an optimized chemical representation of interatomic bonds in the crystal graphs. We demonstrate the accuracy of the improved framework in two distinct illustrations: First, when trained/validated on 180,000/20,000 density functional theory (DFT) calculated thermodynamic stability entries taken from the Open Quantum Materials Database (OQMD) and evaluated on a separate test set of 230,000 entries, iCGCNN achieves a predictive accuracy that is significantly improved, i.e., 20% higher than that of the original CGCNN. Second, when used to assist high-throughput search for materials in the ThCr2Si2 structure-type, iCGCNN exhibited a success rate of 31% which is 310 times higher than an undirected high-throughput search and 2.4 times higher than that of the original CGCNN. Using both CGCNN and iCGCNN, we screened 132,600 compounds with elemental decorations of the ThCr2Si2 prototype crystal structure and identified a total of 97 new unique stable compounds by performing 757 DFT calculations, accelerating the computational time of the high-throughput search by a factor of 130. Our results suggest that the iCGCNN can be used to accelerate high-throughput discoveries of new materials by quickly and accurately identifying crystalline compounds with properties of interest.

💡 Research Summary

The paper presents an enhanced version of the Crystal Graph Convolutional Neural Network (CGCNN), called iCGCNN, designed to improve both predictive accuracy and high‑throughput screening efficiency for crystalline materials. The original CGCNN treats a crystal as a graph where nodes represent atoms and edges represent bonds, but it relies primarily on 2‑body interactions and a relatively simple chemical encoding, which limits its performance on complex, multi‑component systems. iCGCNN addresses these shortcomings through three major innovations.

1. Voronoi‑based structural encoding – By constructing a Voronoi tessellation of the crystal lattice, the model obtains quantitative measures of the spatial partition around each atom, such as face areas, distances, and contact fractions with neighboring atoms. These metrics are appended to node and edge feature vectors, allowing the network to learn physically meaningful proximity information that goes beyond mere Euclidean distances.

2. Explicit three‑body correlations – Real crystals exhibit angular and planar relationships among triplets of atoms that strongly influence electronic structure and bonding. iCGCNN computes the angles and Voronoi face areas for every neighboring atom triplet and incorporates them as additional edge‑wise features. This explicit three‑body term enriches the representation, enabling the network to capture directionality and coordination geometry that are invisible to a pure 2‑body model.

3. Optimized chemical bond representation – In addition to atomic species, the model encodes electronegativity differences, bond order, charge‑transfer propensity, and other chemically relevant descriptors into a high‑dimensional bond vector. Consequently, bonds between the same pair of elements can be distinguished according to their local environment, improving the network’s ability to differentiate subtle variations in bonding.

The architecture retains the multi‑layer graph convolution and pooling scheme of the original CGCNN, but each convolution now processes the expanded feature set. Training is performed on a large subset of the Open Quantum Materials Database (OQMD): 180 000 density‑functional‑theory (DFT) calculated formation energies for training, 20 000 for validation, and a completely independent test set of 230 000 entries. Hyper‑parameters are tuned using Adam optimization, learning‑rate scheduling, and early stopping to avoid over‑fitting. On the test set, iCGCNN reduces the mean absolute error (MAE) by roughly 20 % relative to CGCNN and raises the coefficient of determination (R²) from 0.94 to 0.96, demonstrating a clear gain in predictive fidelity, especially for thermodynamic stability.

To showcase practical impact, the authors apply iCGCNN to a targeted high‑throughput search for new compounds adopting the ThCr₂Si₂ prototype, a structure class known for hosting superconductors, thermoelectrics, and magnetic materials. They generate 132 600 hypothetical elemental decorations of the ThCr₂Si₂ lattice, score each candidate with iCGCNN, and select only the top 5 % (≈6 600) for subsequent DFT verification. After performing 757 DFT calculations, 97 previously unknown stable compounds are identified. This yields a success rate of 31 %, which is 2.4 × higher than the original CGCNN (≈13 %) and 310 × higher than an undirected random high‑throughput search (≈0.1 %). Moreover, the total computational effort is reduced by a factor of about 130 compared with a naïve exhaustive DFT screening.

The study concludes that integrating physically motivated structural descriptors (Voronoi geometry), angular three‑body terms, and richer chemical bond vectors into a graph‑based neural network substantially boosts both accuracy and efficiency for materials discovery. The authors suggest future extensions to predict electronic band structures, transport coefficients, and other functional properties, as well as coupling iCGCNN with active‑learning loops to further accelerate the experimental‑computational feedback cycle. In sum, iCGCNN represents a significant step toward rapid, reliable identification of novel crystalline materials with desired properties.