Source-linked AI summary
Orbital Graph Convolutional Neural Network for Material Property Prediction
Mohammadreza Karamad, Rishikesh Magar, Yuting Shi, Samira Siahrostami, Ian D. Gates, Amir Barati Farimani
TL;DR
Material-property machine learning needs representations that encode local chemistry, yet orbital interactions may be absent or costly to obtain. OGCNN combines orbital and basic atomic features in a crystal graph with an encoder-decoder, and outperforms the compared representations across tested properties.
Problem
Material-property prediction needs representations that encode local chemical environments and orbital interactions, while generating electronic-structure features can add computational cost.
Source: §1 · pp. 1–2Method
OGCNN combines orbital-field and basic atomic features in a crystal graph convolutional network with an encoder-decoder for feature learning.
Source: §1 · p. 2, §2 · pp. 2–3Results
OGCNN has the highest performance among OGCNN, CGCNN, SOAP, and MBTR on the compared properties, except Fermi energy, where it ties SOAP.
Source: §3 · p. 5Takeaways & Limitations
The authors expect OGCNN to be applicable to a broader range of material-discovery applications.
Source: §4 · p. 5
Abstract
from arXiv · showhide
Material representations that are compatible with machine learning models play a key role in developing models that exhibit high accuracy for property prediction. Atomic orbital interactions are one of the important factors that govern the properties of crystalline materials, from which the local chemical environments of atoms is inferred. Therefore, to develop robust machine learningmodels for material properties prediction, it is imperative to include features representing such chemical attributes. Here, we propose the Orbital Graph Convolutional Neural Network (OGCNN), a crystal graph convolutional neural network framework that includes atomic orbital interaction features that learns material properties in a robust way. In addition, we embedded an encoder-decoder network into the OGCNN enabling it to learn important features among basic atomic (elemental features), orbital-orbital interactions, and topological features. We examined the performance of this model on a broad range of crystalline material data to predict different properties. We benchmarked the performance of the OGCNN model with that of: 1) the crystal graph convolutional neural network (CGCNN), 2) other state-of-the-art descriptors for material representations including Many-body Tensor Representation (MBTR) and the Smooth Overlap of Atomic Positions (SOAP), and 3) other conventional regression machine learning algorithms where different crystal featurization methods have been used. We find that OGCNN significantly outperforms them. The OGCNN model with high predictive accuracy can be used to discover new materials among the immense phase and compound spaces of materials
1 Introduction
Machine learning can accelerate material-property prediction, but its accuracy depends on representations that encode local chemical environments and are compatible with the model. OGCNN addresses gaps in prior representations by combining orbital interactions with graph-based learning.
1 Introduction
- High-throughput ab initio calculations can search for materials with desired properties, but their expense motivates faster machine-learning alternatives.Machine-learning models map material representations to properties of interest.Source: §1 · p. 1
- Useful material representations must encode chemical, electronic, geometric, and topological information while remaining distinctive, inexpensive to compute, and compatible with the learning model.Prediction accuracy is largely controlled by how well descriptors encode atoms’ local chemical environments.Source: §1 · p. 1
- Prior descriptors capture elemental or structural information, but orbital interactions are important electronic-structure attributes and are not explicitly represented by some approaches.Generating electronic-structure descriptors can require DFT calculations, increasing computational cost.Source: §1 · pp. 1–2
- CGCNN represents crystal structure, elemental features, and distance-based interactions but omits orbital interactions; OGCNN adds them and is reported to outperform CGCNN and other benchmarked approaches.The benchmarks include MBTR, SOAP, and conventional machine-learning models using different crystal representations.Source: §1 · p. 2
2 Model Architecture
OGCNN combines orbital-field descriptors and basic atomic features in a crystal graph, then uses an encoder-decoder and graph convolutions to predict crystal properties.
2 Model Architecture
- OFM weights each neighbor’s valence-electron representation by a Voronoi solid angle and a distance-dependent function, then concatenates the center atom’s binary vector.For a center atom, the resulting descriptor is a 32 × 33 matrix; Voronoi geometry captures local structure beyond distances alone.Source: §2 · pp. 2–3
- OGCNN represents local orbital interactions with OFM features and combines them with atomic features in a graph-convolutional network.The architecture uses an encoder-decoder to select features before graph convolution and crystal-level prediction.Source: §2 · pp. 2–3
- The OFM input is flattened to 1056 values and combined with 92 basic atomic features to form a 1148-component vector for each atom.Basic atomic features include properties such as group number, period number, and electronegativity.Source: §2 · p. 3
- A multilayer perceptron encoder-decoder transforms the atom features to select important information from the basic atomic and OFM feature pool.The stated layer sizes are 1148 inputs, 768 hidden neurons, and 1536 decoder outputs.Source: §2 · p. 3
- The crystal is an undirected graph with atoms as nodes and bonds as edges; nodes carry combined atomic and OFM features, and multiple edges can connect an atom pair.Kernelized distance features represent the geometry of each bond between neighboring atoms.Source: §2 · p. 3
- Three convolution rounds aggregate neighbor contributions, after which pooling produces a crystal-level representation for property prediction.A two-layer fully connected network maps the pooled output to the target property.Source: §2 · p. 3
3 Training and Results
Across five DFT datasets, OGCNN was evaluated for formation energy, band gap, and Fermi energy, with lower test MAE than CGCNN and the highest performance among the compared representations except for a tie with SOAP on Fermi energy.
3 Training and Results
- The evaluation used five DFT-calculated inorganic-crystal datasets and tested formation energy, band gap, and Fermi energy predictions.Datasets were split 80% for training, 10% for testing, and 10% for validation, with five-fold cross-validation and 100 training epochs.Source: §3 · p. 3
- Figure 2 divides OGCNN into input, encoder-decoder, graph-convolution, and output modules.Source: Figure 2 · p. 4
- OGCNN produced significantly lower test-set MAE than CGCNN across all properties and datasets shown in Figure 3(a).Figure 3(b) reports the percentage reduction in MAE for OGCNN relative to CGCNN.Source: §3 · p. 3, Figure 3 · p. 4
- Formation-energy MAEs were 0.03–0.06 eV/atom across datasets, while OGCNN's band-gap MAE was 0.32 eV versus 0.6 eV reported for DFT calculations.The paper compares these errors with experimental agreement for DFT and desired chemical accuracy for formation energy.Source: §3 · p. 5
- OGCNN had the highest performance among OGCNN, CGCNN, SOAP, and MBTR, except that OGCNN and SOAP tied for Fermi energy.The comparison covered five datasets and different material properties.Source: Table 1 · p. 5, §3 · p. 5
4 Conclusion
The conclusion presents OGCNN as a model combining atomic orbital interactions with basic atomic features and reports higher accuracy than previously reported models. The authors expect it to apply to a broader range of material-discovery applications.
4 Conclusion
- OGCNN embeds atomic orbital-orbital interaction features and basic atomic features, and was applied to predict diverse material properties across versatile structures.Source: §4 · p. 5
- The authors report significantly higher prediction accuracy than previously reported models, attributing superior performance to orbital interactions and an encoder-decoder network.Source: §4 · p. 5
- The authors expect OGCNN to apply to a broader range of material-discovery applications.The paper provides the repository at Ref..Source: §4 · p. 5