Prediction of material parameters of spin textures from LTEM images by machine learning models CNN and ViT

Machine learning (ML) models were trained to predict the Dzyaloshinskii–Moriya interaction constant D and anisotropy constant K of spin textures from Lorentz transmission electron microscopy (LTEM) images. Two ML models, convolutional neural network (CNN) and vision transformer (ViT), were trained, tested and employed to predict the values of D and K. Firstly, training and testing datasets composed of 9300 topological spin textures were prepared using the micromagnetic simulation method with the values of D and K randomly chosen. Secondly, the performance of the CNN model for predicting D and K values was investigated by varying the number of training data, the pooling process, and the number of fully connected layers, which indicated high prediction accuracies. Thirdly, better performance of the ViT model for predicting D and K values was achieved with the coefficient of determination R² reaching as high as 0.9987 and 0.9991 for predicting D and K values, respectively. Particularly, the reliance on synthetic data and noise robustness was discussed. Finally, the prediction abilities of CNN and ViT models were evaluated and compared. The results of this research indicate that the ML models can achieve the material parameters directly and effectively from LTEM images, and this process may contribute to the design of advanced devices based on the topological spin textures.