Machine-Learning Angle-Resolved Photoemission Spectroscopy under Tunable Magnetic Fields
PI: Chaoxing Liu
Investigators
PI: Prof. Chaoxing Liu – Physics Department. Co-PI: Prof. Heike Pfau and Prof. Dezhe Jin – Physics Department. Liu, who serves as the main mentor, is a condensed matter theorist focusing on electronic properties of quantum materials. Jin is a computational physicist with expertise in training artificial neural network models for many-body quantum systems and simulating neuronal dynamics that underline brain functions. Pfau is leading the experimental side of this project, and her group performs magneto-ARPES measurements on various quantum materials.
Project Information (See project description in the last page of this document)
Project Title: Machine-Learning Angle-Resolved Photoemission Spectroscopy under Tunable Magnetic Fields Research areas: Quantum physics and materials, Computational science, AI. Relevant ICDS Center: CENSAI: The Center for Artificial Intelligence Foundations and Scientific Applications. Connection with ICDS’ mission: This project closely aligns with the objectives of CENSAI by leveraging machine learning approach to guide the design of cutting-edge experiments and drive foundational advances in quantum materials research.
Key objectives
Our project consists of two main components, each of which is publishable independently or in combination with experimental results. 1. Optimizing magnetic field configuration in the ARPES apparatus to minimize magnetic field distortion effect on electron trajectories by utilizing machine learning tools such as PyTorch. 2. Generating a large theoretical simulation dataset for magneto-ARPES images, and training AI models (such as neural networks) to reconstruct the electronic band structures of materials from magneto-ARPES data.
Medium/long-term goals (e.g., a successful proposal to a specific call)
The seed proposal aims to lay the theoretical foundation for developing a new experimental technique of magneto-ARPES. Co-PI Pfau previously had a 2-year grant “Momentum-resolved Majorana Band Spectroscopy” from DOE, aiming to develop experimental techniques of magnetoARPES, and her group will continue this effort. Thus, we anticipate that this seed proposal, which primarily focuses on the theoretical and numerical aspects of this technology, will naturally pave the way for a successful experiment-theory collaboration proposal aimed at applying this technology to the exploration of magnetic topological materials and topological superconductors for quantum computing in the medium term. In the long term, we aim to advance the technology of magneto-ARPES to operate with magnetic fields beyond sub-Tesla range, enabling the observation of Landau level physics in quantum materials. This will have a huge impact in condensed matter physics community and significantly deepen our understanding of exotic quantum states, such as integer and fractional quantum Hall effects.
Expertise from Junior Researcher
To support the proposed research, we seek to collaborate with an ICDS junior researcher whose expertise bridges physics, computational methods, and machine learning. Specifically 1. A strong background in physics, ideally with familiarity in angle-resolved photoemission spectroscopy (ARPES) and quantum materials; 2. Proficiency in computational physics, including the numerical solution of differential equations relevant to modeling electron trajectories under magnetic fields; 3. Experience with high-performance computing (HPC) environments and parallel programming techniques, e.g. using OpenMP with C/C++ or python; 4. Working knowledge of machine learning frameworks such as TensorFlow or PyTorch, with an emphasis on applying machine learning approach to physical dataset; 5. The ideal candidate would be a post-comprehensive-exam Ph.D student.
Level of effort
We estimate this research project requires 2 semesters of 50% RA. The remaining tuition funding for the graduate student is covered by PI’s other federal funds, such as NSF grant, or from TA, depending on the funding situation. Engagement with ICDS Liu’s research group has been actively utilizing the ICDS Roar computing facility to support graduate students’ research in nonlinear quantum transport. These computational efforts directly complement experimental investigations within the Physics Department. Access to Roar has significantly enhanced our ability to perform complex calculations efficiently and on a timely basis, enabling faster iteration between theory and experiment. Jin has been the early adopter of ICDS’s credit system for computing and tested the early version of this system. He has extensively used the GPUs for the quantum wave function project this semester and also provided suggestions for the future GPU purchasing plans. This project offers a clear and immediate opportunity to engage with the ICDS mission by applying advanced computational and machine learning methods to a cutting-edge problem in quantum material research. In the short term, we plan to engage with ICDS through consultation with its technical experts to support the development of neural network models in understanding new technology of magneto-ARPES. We will utilize ICDS’s high-performance computing resources for data generation, model training and optimization using PyTorch, enabling scalable and efficient development. This collaboration will foster valuable synergy between theory, computation, and experiment, exemplifying the interdisciplinary integration at the core of ICDS’s vision.
Background, Motivation and Opportunity: Angle-resolved photoemission spectroscopy (ARPES) is an experimental technique that directly probes electronic band structures in momentum space and has significantly advanced our understanding of the fundamental electronic properties of quantum materials over the past two decades (RMP, 93, 025006, 2021). However, ARPES cannot probe materials under magnetic fields, thereby restricting its applicability for investigating magnetic quantum materials. This is because magnetic fields exert a Lorentz force on electrons, distorting their trajectory and preventing accurate measurements of their momentum immediately after photoemitted from the materials. Recently, there have been efforts at prototypical level to understand the ARPES band structures with in-situ magnetic fields, a technique called “magneto-ARPES” (JESRP, 266, 147357, 2023; RSI, 94, 093902, 2023). However, due to the complexity in electron trajectory caused by non-uniform magnetic fields, it remains a great challenge to extract band structure information from ARPES images using magneto-ARPES. Overcoming this obstacle will enable the concurrent tuning of magnetic states through applied fields and measurement of band structures, thereby opening a new avenue in exploring magnetic quantum materials. Co-PI Pfau’s group at PSU has recently developed the experimental technique of magneto-ARPES and observed high-quality ARPES data for quantum materials under magnetic fields. Thus, theoretical simulations and the interpretation of magneto-ARPES data are highly demanding. Recent advancements in artificial intelligence, particularly machine learning (ML), offer powerful tools for tackling challenging problems that lie beyond the reach of traditional approaches. Although solving electron trajectories in non-uniform magnetic fields is a textbook-level problem, optimizing non-uniform magnetic field configurations and simulating ARPES image of band structures under realistic experimental conditions (e.g. finite sample size) are more challenging, where ML approach can play a crucial role.
Proposed Work: The objective of this proposal is to leverage the ML approach to (1) design magnetic field configurations that minimize the distortions of electron trajectories, and (2) develop a neural network model to extract band structure from ARPES images under tunable magnetic fields. Our first task is to develop a numerical approach for simulating magneto-ARPES images under realistic experimental conditions with non-uniform magnetic fields that are generated by coils surrounding the samples. Coil configurations will be optimized to reduce the influence of non-uniform magnetic fields on electron trajectories, which will be done using ML tools such as PyTorch. Numerical simulations will be compared with experimental ARPES data to identify key ingredients that improve the momentum resolution of ARPES images. After developing the numerical approach, we will implement numerical simulations of various well-known band structures and collect the simulation data for ARPES images with varying magnetic fields. Half of simulation data will be utilized to train a neural network (NN) model to revert the distorted ARPES images back to restore the band structure of materials without magnetic field distortion, while the other half will be used to test and validate this NN model. This NN model will be further calibrated using experimental data with the aim of extracting electron band structures from ARPES images experimentally measured under magnetic fields. The feedback from experiments will further improve our numerical simulation model. This iterative approach, combining theoretical simulations, ML-guided design and experimentation, will significantly enhance our understanding of magnetic field distortion effect on ARPES spectra and pave the way for future investigations of quantum materials such as topological superconductors for quantum computing using magneto-ARPES.
Key personnel: Liu is a condensed matter theorist focusing on electronic properties of quantum materials. Jin is a computational physicist with expertise in training artificial NN models for many-body quantum systems and simulating neuronal dynamics that underline brain functions. Pfau is leading the experimental side of this project, and her group performs magneto-ARPES measurements on various quantum materials.