Research And Application Of Crystal Property Prediction Based On Deep Learning

Nowadays,the discovery and application of new crystal materials is crucial to the progress of science and the development of civilization,and the emergence of density functional theory has provided a practical calculation method for the study of crystal properties.With the development and application of supercomputers in recent years,researchers can use its powerful computing capabilities to perform accurate density functional theory calculations.Meanwhile,high-throughput calculations based on density functional theory can generate large amounts of data,making it possible to apply machine learning methods based on data-driven in the field of materials science.Unlike traditional research methods in materials science,machine learning processes much larger amounts of data and extracts information with a greater degree of automation.As a subfield of machine learning,deep learning has grown rapidly in recent years.Through deep neural networks composed of multiple processing layers,deep learning can transform low-level features into more complex and abstract high-level features and discover inherent patterns and regularities from large-scale datasets.Owing to its excellent performance,deep learning has been gradually applied by researchers to the field of materials science.For the application of deep learning in the prediction of crystal properties,we have carried out the following work:1.Combining attention mechanism and deep learning-based crystal graph convolutional neural network,an attention mechanism-based crystal graph convolutional neural network is proposed.The attention mechanism is used to learn the importance of all neighbours and assign different weights to different neighbours according to their corresponding importance,so that the network model can screen out the neighbours with strong relevance and ignore those with weak relevance.This enhances the model’s ability to fuse crystal topology and atomic features,and improves the accuracy of prediction.The node normalisation method is also introduced to regularizes the model by suppressing the feature correlation of the embedding and increasing the model smoothing associated with the input node features.The risk of overfitting in the network is effectively reduced and the training process is stabilised.The experimental results show that the prediction accuracy of the attention mechanism-based crystal graph convolutional neural network for total energy,bandgap,Fermi energy and formation energy is higher than that of the original crystal graph convolutional neural network.2.The residual learning is introduced into the crystal graph convolutional neural network,and a crystal graph convolutional residual neural network is proposed.The crystal graph convolutional residual neural network adds a fully connected residual neural network after the pooling layer,which uses the.residual connection method for each fully connected layer.Then adds batch normalization to avoid network overfitting due to increased model complexity.Moreover,ELU activation function is introduced to perform non-linear transformations to enhance the expressiveness and learning ability of the model.The experimental results show that the crystal graph convolutional residual neural network has higher prediction accuracy for total energy,bandgap,Fermi energy and formation energy than the original crystal graph convolutional neural network.3.Under the framework of the crystal graph convolutional neural network and the attention mechanism-based crystal graph convolutional neural network,the wide bandgap semiconductors classification experiments with the bandgap threshold of 2.3 eV and the ferromagnetic crystals classification experiments with the total magnetization threshold of 0.5μB were performed using the negative log-likelihood loss function.The experimental results show that the crystal graph convolutional neural network and the attention mechanism-based crystal graph convolutional neural network can achieve high-precision wide bandgap semiconductors classification and ferromagnetic crystals classification.At the same time,compared with the crystal graph convolutional neural network,the attention mechanism-based crystal graph convolutional neural network achieves higher classification accuracy under the condition of the same amount of training data,and can train a model with the same classification accuracy using less data.