Source-linked AI summary
Developing an improved Crystal Graph Convolutional Neural Network framework for accelerated materials discovery
Cheol Woo Park, Chris Wolverton
TL;DR
DFT-based discovery remains computationally challenging because of the sheer size of the materials search space. The paper presents iCGCNN as an improved CGCNN framework and reports better performance, 97 new stable compounds, and a 130-fold accelerated search.
Problem
DFT-based discovery of novel materials remains computationally challenging because of the sheer size of the materials search space.
Method
The paper presents an improved CGCNN model, using crystal-structure information including Voronoi construction.
Results
20% better predictive accuracy and a 2.4-fold higher success rate than CGCNN were reported, alongside 97 new stable compounds and a 130-fold accelerated high-throughput search.
Takeaways & Limitations
The results suggest that iCGCNN can accelerate high-throughput discovery of new materials.
Takeaways & Limitations
The training data include unstable compounds, and incomplete convex hulls constrain the available stability knowledge.
Abstract
from arXiv · showhide
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.
Introduction
Materials discovery with DFT is computationally challenging because of the large search space, motivating machine-learning models that learn useful crystal representations. This work introduces iCGCNN, an improved CGCNN framework, and demonstrates stronger performance in stability prediction and high-throughput discovery.
- DFT-based discovery of novel materials remains computationally challenging because the materials search space is very large.
- iCGCNN improves CGCNN by incorporating Voronoi-tessellated structure information, explicit 3-body correlations, and optimized interatomic-bond representations.These descriptors are absent from the original crystal graphs.
- The study evaluates CGCNN and iCGCNN on DFT-calculated thermodynamic stability data from the Open Quantum Materials Database.The models use either predicted formation energies followed by convex-hull calculations or direct thermodynamic-stability prediction.
- iCGCNN achieves thermodynamic-stability prediction accuracy that is 20% higher than CGCNN.
- In ThCr2Si2-type high-throughput searches, iCGCNN identifies nearly twice as many stable compounds and achieves a success rate 2.4 times higher than CGCNN.Using both models, the study screened 132,600 compounds and identified 97 stable compounds with 757 DFT calculations, 130 times faster than an undirected search.
Results
The iCGCNN improves crystal-graph representations with structural and chemical information, yielding lower stability-prediction errors and more efficient discovery of stable ThCr2Si2-type compounds than CGCNN.
- Description of improvements in the iCGCNN: iCGCNN augments crystal graphs with Voronoi-tessellated structure information, explicit 3-body correlations, and an optimized interatomic-bond representation.Its graph embeds atom features and neighbor correlations, while edges include Voronoi-polyhedron descriptors such as solid angle, area, and volume.
- Results: 30.5 meV/atom formation-energy MAE was achieved by iCGCNN versus 41.3 meV/atom for CGCNN.Hull-distance errors were correspondingly 30 and 41 meV/atom for iCGCNN and CGCNN, respectively.
- Using CGCNN and iCGCNN to accelerate high-throughput DFT searches: 757 DFT calculations identified 143 stable unique compounds, including 97 not previously reported in the literature; the search was accelerated by a factor of 130.The 97 compounds remain computational predictions awaiting experimental validation.
Conclusions
The paper presents iCGCNN as an improved CGCNN framework by enriching crystal representations and evaluates it on hull-distance prediction and ThCr2Si2 materials discovery. iCGCNN outperforms CGCNN in both predictive accuracy and targeted screening, while the combined workflow identifies new stable compounds efficiently.
- iCGCNN integrates Voronoi tessellation information, explicit 3-body correlations, and an optimized chemical representation of interatomic bonds.
- 2.4 times greater success rate and nearly twice as many stable compounds were obtained by iCGCNN than by CGCNN for ThCr2Si2-type materials.
- 97 new stable compounds were discovered using 757 DFT calculations, accelerating the high-throughput search by a factor of 130.
- The results support iCGCNN as an efficient screening tool for predicting potentially stable materials and accelerating materials discovery.
Data
The study uses DFT-calculated thermodynamic data from OQMD to train, validate, and test its machine-learning models. OQMD combines experimentally known and hypothetical inorganic compounds calculated with VASP, while the test set is separate from training and validation data.
- DFT-calculated thermodynamic data from OQMD are used for training, validation, and testing throughout the work.
- OQMD v1.1 contains about 450,000 DFT-calculated unique ordered inorganic compounds, including about 40,000 experimentally known compounds.
- The remaining OQMD compounds are hypothetical structures generated from commonly occurring structural prototypes.
- All OQMD DFT calculations are performed with the Vienna Ab Initio Simulation Package, VASP.
- Models are trained on about 180,000 compounds, validated on about 20,000, and tested on a separate set of about 230,000 compounds.