.png){kind=icon}
.png){kind=icon}
Return
Physical encoding improves OOD performance in deep learning materials property prediction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.commatsci.2024.113603.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep learning (DL) models have been widely used in materials property prediction with great success, especially for properties with large datasets. However, the out-of-distribution (OOD) performances of such models are questionable, especially when the training set is not large enough. Here we evaluated four atomic encoding methods for predicting six material properties. Our comprehensive experiments showed that using physical (atomic) encoding rather than the widely used one-hot encoding for atoms/elements can significantly improve the OOD performance by increasing the models' generalization performance, which is especially true for models trained with small datasets. Our benchmark results of both composition- and structure-based deep learning models over six datasets including formation energy, band gap, refractive index, and elastic properties predictions demonstrated the importance of physical atomic encoding to OOD generalization for models trained on small datasets.
Keywords:
Material property prediction
Out-of-distribution
Machine learning
Graph neural networks
Physical Encoding
One-hot encoding
Journal
IF:
3.3
Papers:
1.3W
Citations:
3.6W