Machine Learning Reconstruction of High-Dimensional Electronic Structure from Angle-Resolved Photoemission Spectroscopy
Yu Zhang, Yong Zhong, Nhat Huy Tran, Shuyi Li, Kyuho Lee, Yonghun Lee, Tiffany C. Wang, Harold Y. Hwang, Zhi-Xun Shen, Chunjing Jia · 2026
Read Machine Learning Reconstruction of High-Dimensional Electronic Structure from Angle-Resolved Photoemission Spectroscopy by Yu Zhang, Yong Zhong, Nhat Huy Tran, Shuyi Li, Kyuho Lee, Yonghun Lee, Tiffany C. Wang, Harold Y. Hwang, Zhi-Xun Shen, Chunjing Jia (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-16725.txt.
Introduction
Zhang and co-authors use deep learning to read electronic structure from photoemission spectra. Such spectra map energy and momentum across four dimensions. Getting model parameters out of them is slow work done by hand. Their neural net speeds that step up a great deal. They test it on two oxide families, nickelates and manganites. The fits beat classic by-hand fitting on Fermi surfaces and band shapes. Labs could run the same checks the same way. The team aims at self-running discovery pipelines.
Catalog record
| CREATOR | Yu Zhang, Yong Zhong, Nhat Huy Tran, Shuyi Li, Kyuho Lee, Yonghun Lee, Tiffany C. Wang, Harold Y. Hwang, Zhi-Xun Shen, Chunjing Jia |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 23276 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.16725
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Machine Learning Reconstruction of High-Dimensional Electronic Structure from Angle-Resolved Photoemission Spectroscopy Yu Zhang, Yong Zhong, Nhat Huy Tran, Shuyi Li, Kyuho Lee, Yonghun Lee, Tiffany C. Wang, Harold Y. Hwang, Zhi-Xun Shen, Chunjing Jia Abstract The emergent behavior of quantum materials is governed by their electronic structure, which can be experimentally probed by photoemission spectroscopy techniques that generate a four-dimensional dataset of energy and momentum. However, the quantitative extraction of Hamiltonian parameters from these high-dimensional spectra remains a significant challenge, currently relying on labor-intensive, expert-dependent analysis rather than standardized workflows. Here, we introduce a deep learning framework based on implicit neural representations to accelerate the retrieval of Hamiltonian parameters in two types of transition-metal oxides: perovskite nickelates and manganites. Our approach outperforms traditional analytical fitting procedures, yielding superior agreement with experimental Fermi surface topologies and energy-momentum dispersions. This work highlights the potential of deep learning tools to bridge the gap between theory and experiment, paving the way for high-throughput, autonomous discovery pipelines in quantum materials. Introduction Artificial Intelligence (AI) has catalyzed a paradigm shift in the natural sciences, most dramatically in biology, where deep learning architectures such as AlphaFold have solved long-standing challenges in protein structure prediction and gene sequence analysis 12 ; 11 . This breakthrough provides a compelling blueprint for the next frontier in materials science: the AI-driven discovery of quantum materials. By leveraging deep learning tools to decode materials’ “genome” — the underlying electronic structures, novel quantum states can be rationally designed to meet the demands of high-temperature superconductivity, exotic magnetism and non-trivial topology. Modern AI techniques have emerged as a transformative force in the theoretical prediction of electronic structures. At the density functional theory (DFT) level, recent developments include efficient mapping of potential energy surfaces via DeepMD 34 ; 33 and general machine learning interatomic potentials (MLIPs) such as EquiformerV2 and MatterSim 18 ; 38 , as well as direct prediction of electronic band dispersions using frameworks like DeepH 16 ; 9 . These pioneering approaches learn from large datasets of first-principles DFT or ab initio molecular dynamics calculations, enabling the simulation for greatly more complex systems with remarkable efficiency. Beyond interatomic potentials, neural-network quantum states (NQS) provide a powerful route to solving strongly correlated systems based on model Hamiltonians. By employing restricted Boltzmann machine (RBM) 2 and Transformers-based 35 ; 28 architectures, open-source tools such as NetKet 30 and DeepSolid 17 capture strong electronic correlations and entanglement effects beyond the mean-field level. Recent developments have extended these capabilities to spectral functions 21 ; 32 ; 10 , enabling investigation of both static and dynamical electronic properties. Although rapid advances have been achieved in the theoretical frontier of electronic structure prediction, the ability to benchmark these results with experimental observations remains a critical bottleneck. Angle-resolved photoemission spectroscopy (ARPES) is the most direct experimental technique for probing the electronic structure of quantum materials 27 . By resolving both energy ( E ) and momentum ( \mathbf{k} ), it generates a four-dimensional dataset ( E,k_{x},k_{y},k_{z} ) that encapsulates the full electronic information. Despite these unparalleled capabilities, high-quality ARPES datasets are severely limited by practical barriers: experiments require access to large-scale synchrotron facilities and involve low-throughput acquisition cycles spanning tens of hours. Moreover, downstream analysis is labor-intensive and contingent upon expert bias. Therefore, the development of dedicated AI frameworks is essential to accelerate the analysis of high-dimensional spectral data. Deep learning has demonstrated increasingly important utilities across diverse spectroscopic techniques. It has been widely used to uncover crucial material properties through feature extraction 6 ; 7 ; 37 , high-fidelity denoising 25 and spectral prediction 1 . Within the specific domain of ARPES, nascent efforts have employed deep learning architectures to address specialized problems, such as probabilistic band retrieval 31 , superconducting gap determination 5 , and multidimensional data cleaning 13 . Despite these localized successes, a generalized AI framework capable of directly extracting fundamental electronic parameters from raw ARPES data remains elusive. This objective is hindered by several challenges: the scarcity of standardized experimental repositories, the intrinsic complexity of many-body electronic interactions, and the experimental noise inherent to the measurement process. In this work, we address this critical gap by providing a first-of-its-kind approach to capturing fundamental electronic Hamiltonian parameters directly from experimental ARPES data. We incorporate a physics model to guide a machine learning (ML) architecture in learning the underlying physics. We use two real-world examples, perovskite nickelate and manganite, to demonstrate the power of deep learning methods for ARPES data interpretation. The algorithm delivers reconstructions that agree excellently with experimental spectra for both materials, demonstrating that an ML‑based approach can automatically extract high‑dimensional electronic parameters with markedly greater clarity, computational efficiency, and transferability than conventional techniques. Results The perovskite nickelate Nd 1-x Sr x NiO 3 is a promising candidate for next-generation electronic devices, characterized by a sharp, tunable metal-insulator transition 19 ; 29 ; 3 ; 22 ; 4 . Its low-energy electronic structure is primarily governed by the two e_{g} orbitals of the \mathrm{Ni}^{3+} ion, as shown in Figure 1 a . Recently, systematic ARPES measurements have succeeded in mapping the high-quality, four-dimensional electronic structure of Nd 1-x Sr x NiO 3 with x=0.175 36 . This rich experimental dataset makes perovskite nickelate an ideal model system for developing ML tools to automatically extract the essential electronic parameters underpinning its unique properties. Figure 1: a . Crystal structure of Nd 1-x Sr x NiO 3 with the corresponding electronic configuration of Ni 3+ . b . Architecture of the SIREN model: five hidden layers are placed between the inputs (three tight-binding parameters and \mathbf{k} ) and the output (spectral intensity I_{ML} ). c . Fermi surface obtained from the trained machine learning model (left) and from the tight-binding model (right). Two-Band Tight-Binding Model To capture the essential low-energy physics of perovskite nickelate, we employ a tight-binding model to simulate its electronic structure. This model is designed to reproduce two critical features in the ARPES study: band dispersion and Fermi surface topology. The band dispersion, E(\mathbf{k}) , describes the relationship between the energy of an electron and its momentum. The Fermi surface—the boundary in momentum space separating occupied from unoccupied states—dictates many of the key properties of the material. Based on the electronic configuration of Ni 3+ ion, a two-band tight-binding Hamiltonian can provide an effective description of the band dispersion and the Fermi surface topology for Nd 1-x Sr x NiO 3 system 15 : H_{tb}=-\sum_{ij}t_{ij}^{ab}c_{ia}^{\dagger}c_{jb}-\mu (1) Here, i , j are site indices, a,b=1,2 are orbital indices corresponding to d_{3z^{2}-r^{2}} and d_{x^{2}-y^{2}} orbitals, and \mu is the chemical potential which can be tuned by changing the doping levels. Only the nearest-neighbor hopping t_{1} and next-nearest-neighbor hopping t_{2} with \sigma -type bonding are considered in our simulations. We can calculate ARPES spectral intensity I_{TB}(E,\mathbb{k}) through equations (2) and (3), where \epsilon(\mathbf{k}) represents the bare band, \Sigma^{\prime}(E) and \Sigma^{\prime\prime}(E) are the real and imaginary parts of self-energy, respectively, and f(E) is the Fermi-Dirac function. The k_{z} broadening is taken into account by applying a Lorentzian convolution with \Delta k_{z}=0.2\frac{\pi}{a} . The resulting spectral intensity varies with different choices of (t_{1},t_{2},\mu) . A(E,\mathbb{k})=\sum-\frac{1}{\pi}\frac{\Sigma^{{}^{\prime\prime}}(E)}{(E-\epsilon(\mathbb{k})-\Sigma^{{}^{\prime}}(E))^{2}+(\Sigma^{{}^{\prime\prime}}(E))^{2}} (2) I_{TB}(E,k_{x},k_{y})=\int_{k_{z}^{0}-\frac{\Delta k_{z}}{2}}^{k_{z}^{0}+\frac{\Delta k_{z}}{2}}f(E)\frac{A(E,\mathbb{k})}{(k_{z}-k_{z}^{0})^{2}+(\Delta k_{z}/2)^{2}} (3) Implicit Neural Representations Conventional analytical approaches for