Prediction Of Conductivity Properties Of Materials Based On Machine Learning

Spin Hall Conductance（SHC）and Anomalous Hall Conductance（AHC）of materials are important properties of conductor materials.Different from the ordinary Hall effect,the anomalous Hall effect is generated by the spontaneous magnetization of the material,and materials with higher Anomalous Hall Conductance It can be used to make lowenergy devices and reduce the heating of computers during operation.The spin Hall effect is caused by the different spin directions of electrons in materials.Materials with high Spin Hall Conductance can be used in quantum computers.High conductance chip.In the field of materials science,high-throughput methods are usually used to calculate Spin Hall Conductance and Anomalous Hall Conductance,which takes a long time.The general calculation cycle is one week.The two have a certain influence.Therefore,this thesis uses the method of machine learning to predict the two based on the energy band data,one is to improve the calculation efficiency,and the other is to provide a certain reference for the theoretical research on the specific impact of the energy band data on the two.In this thesis,two models are used to predict the Spin Hall Conductance and Anomalous Hall Conductance,namely the Light-GBM model and the deep neural network model in the ensemble learning in the machine learning method.The difference between the two models lies in the way of feature extraction.The Light-GBM model uses the traditional method to extract the slope of the energy band,the signal feature and the degree of energy band splitting;when building the neural network model,the commonly used method 1D-CNN+ Starting from LSTM,a model combining Res Net50 module,BiLSTM module and Transformer feature extraction module using one-dimensional convolution is proposed,and energy band features,time series features and inter-channel features are extracted respectively,and a band splitting method is proposed.Approach.Experiments were carried out on two models using four data sets.The experimental content of Light-GBM is mainly control experiments,and the models selected for the control experiments are SVR and XGBoost;the experimental content of the neural network model includes ablation experiments and control experiments,ablation experiments.The purpose is to verify the importance of the Transformer module.The model selected for the control experiment is 1D-CNN+LSTM as the baseline,and the Res Net50 model is a different feature extraction method.Four evaluation indicators,mean square error,root mean square error,mean absolute error and coefficient of determination are selected,and the training time is counted.Finally,the experimental results are compared and analyzed from the above perspectives.The experiments done in this thesis prove that the feature extraction method adopted by Light-GBM performs well,and the neural network model has improved the prediction accuracy by more than 50% compared with the baseline,which also proves the importance of the Transformer module in this model.The prediction effect of energy band data on SHC is better than that of AHC.At the same time,the training time of Soc energy band data is longer.The average training time of Light-GBM is 174 s,and the average root mean square error is 15.97.The training time of neural network model The time averaged 5.16 hours and the RMSE averaged 7.64.