Machine Learning-Augmented Spectroscopies for Intelligent Materials Design (Springer Theses)

Description

The thesis contains several pioneering results at the intersection of state-of-the-art materials characterization techniques and machine learning. The use of machine learning empowers the information extraction capability of neutron and photon spectroscopies. In particular, new knowledge and new physics insights to aid spectroscopic analysis may hold great promise for next-generation quantum technology. As a prominent example, the so-called proximity effect at topological material interfaces promises to enable spintronics without energy dissipation and quantum computing with fault tolerance, yet the characteristic spectral features to identify the proximity effect have long been elusive. The work presented within permits a fine resolution of its spectroscopic features and a determination of the proximity effect which could aid further experiments with improved interpretability. A few novel machine learning architectures are proposed in this thesis work which leverage the case when the data is scarce and utilize the internal symmetry of the system to improve the training quality. The work sheds light on future pathways to apply machine learning to augment experiments. Supervisor's Foreword Acknowledgments Parts of This Thesis Have Been Published in the Following Journal Articles and Preprints Contents 1 Introduction 1.1 Neutron and Photon Scattering and Spectroscopy 1.2 Integration of Machine Learning 1.3 Thesis Objectives References 2 Background 2.1 Neutron and Photon Scattering and Spectroscopy 2.1.1 Inelastic Neutron Scattering 2.1.2 Raman Spectroscopy 2.1.3 Polarized Neutron Reflectometry 2.1.4 X-ray Absorption Spectroscopy 2.2 Data-Driven Methods 2.2.1 Dimensionality Reduction Singular Value Decomposition Principal Component Analysis Non-negative Matrix Factorization 2.2.2 Machine Learning Support Vector Machines Neural Networks References 3 Data-Efficient Learning of Materials' Vibrational Properties 3.1 Introduction 3.2 Materials Data Representations 3.3 Euclidean Neural Networks 3.3.1 Graph Representation of Crystal Structures 3.3.2 Network Operations 3.4 Phonon DoS Prediction 3.4.1 Data Processing 3.4.2 Results 3.4.3 Comparison with Experiment 3.4.4 High-CV Materials Discovery 3.4.5 Partial Phonon Density of States 3.4.6 Alloys and Strained Compounds 3.5 Unsupervised Representation Learning of Vibrational Spectra 3.5.1 Dimensionality Reduction 3.5.2 Data Processing Methods 3.5.3 Results 3.6 Conclusion References 4 Machine Learning-Assisted Parameter Retrieval from Polarized Neutron Reflectometry Measurements 4.1 Introduction 4.2 Polarized Neutron Reflectometry 4.3 Variational Autoencoder 4.3.1 VAE-Based PNR Parameter Retrieval 4.3.2 Data Preparation 4.3.3 Results 4.4 Resolving Interfacial AFM Coupling 4.5 Discussion 4.6 Conclusion References 5 Machine Learning Spectral Indicators of Topology 5.1 Introduction 5.2 Topological Materials Discovery 5.3 Data Preparation and Pre-processing 5.4 Exploratory Analysis 5.5 Results 5.6 Conclusion References 6 Conclusion and Outlook 6.1 Thesis Summary 6.2 Perspectives and Outlook Reference