卷积神经网络辅助无机晶体弹性性质预测

Inorganic crystal materials have shown extensive application potential in many fields due to their excellent physical and chemical properties. Elastic properties, such as shear modulus and bulk modulus, play an important role in predicting the electrical conductivity, thermal conductivity and mechanical properties of materials. However, the traditional experimental measurement method has some problems such as high cost and low efficiency. With the development of computational methods, theoretical simulation has gradually become an effective alternative to experiments. In recent years, graph neural network-based machine learning methods have achieved remarkable results in predicting the elastic properties of inorganic crystal materials, especially, crystal graph convolutional neural networks (CGCNNs), which perform well in the prediction and expansion of material data. In this study, two CGCNN models are trained by using the shear modulus and bulk modulus data of 10987 materials collected in the Matbench v0.1 dataset. These models show high accuracy and good generalization ability in predicting shear modulus and bulk modulus. The mean absolute error (MAE) is less than 13 and the coefficient of determination (R² ) is close to 1. Then, two datasets are screened for materials with a band gap between 0.1 and 3.0 eV and the compounds containing radioactive elements are excluded. The dataset consists of two parts: the first part is composed of 54359 crystal structures selected from the Materials Project database, which constitute the MPED dataset; the second part is the 26305 crystal structures discovered by Merchant et al. (2023 Nature 624 80) through deep learning and graph neural network methods, which constitute the NED dataset. Finally, the shear modulus and bulk modulus of 80664 inorganic crystals are predicted in this study This work enriches the existing material elastic data resources and provides more data support for material design.