PaperlayerView exact revision on arXiv

Source-linked AI summary

Progress and Prospect of AI in ARPES Workflow

Sandy Adhitia Ekahana, Aalok Tiwari, Pratik Saud, Aaron Bostwick, Chris Jozwiak, Eli Rotenberg, Jyoti Katoch

arXiv:2610.10140v1cond-mat.othercs.LGphysics.data-an

TL;DR

ARPES research faces increasingly complex multidimensional measurements, while a comprehensive review of AI and ML applications and their reliability across the workflow has been lacking. This paper surveys methods and applications across that workflow and the available data landscape, concluding with proposals for pretrained task models and a more standardized open archive.

  • Problem

    ARPES measurements span many dimensions, and a comprehensive review of AI and ML applications and reliability across the workflow is lacking.

    Source: §2 · p. 3, §1 · p. 3

  • Method

    The paper reviews relevant AI and ML methods and applications across the ARPES workflow, data landscape, and prospects for laboratory automation.

    Source: §1 · p. 3

  • Results

    The review proposes sharing pretrained task models alongside work toward a larger, standardized open ARPES dataset repository.

    Source: §1 · p. 3, §9 · p. 24

  • Takeaways & Limitations

    A federated open archive of task-specific benchmarks and models is proposed as a lower-inertia route to AI adoption in the ARPES pipeline, not a substitute for larger data deposits.

    Source: §9 · p. 24

    Long-acquisition maps are not automatically perfect ground truth because surface aging, beam drift, charging, and photon-flux changes can alter measurement states.

    Source: §9 · p. 23

Abstract

from arXiv · show

Artificial intelligence (AI) is becoming an increasingly useful tool across the experimental sciences, including angle-resolved photoemission spectroscopy (ARPES), which routinely produces large, multidimensional datasets of electronic structure. Recent advances in AI and machine learning (ML) have opened new opportunities across the entire ARPES workflow, from automated sample preparation and real-time data acquisition to post-experiment data analysis and comparison with theoretical calculations. Despite this progress, a comprehensive review of ML applications, their capabilities, and reliability across the different stages of ARPES workflow is still lacking. In this review, we first introduce ML methods that are most relevant to experimentalists working in condensed matter physics and materials science. We then follow the ARPES workflow, reviewing existing ML applications at each step and discussing their advantages, limitations and potential for future development. We also examine the current ARPES data landscape, where several open databases are available but remain relatively small and fragmented compared with large, shared datasets such as ImageNet. Given these limitations, we suggest that the community focus on sharing pretrained models that can be further trained, adapted to specific tasks, and redistributed, while working toward a larger and standardized open ARPES dataset repository. Finally, we discuss our perspectives on the future of AI within the ARPES workflow using a six-level framework of laboratory automation, highlighting the opportunities and challenges in moving toward a fully autonomous, self-driving ARPES laboratory.

1 Introduction

AI and ML are increasingly practical across scientific research, with applications in condensed matter physics and experimental workflows. This review surveys AI and ML across the full ARPES workflow, from sample preparation and measurement through analysis, data resources, and future automation.

1 Introduction

  • Accessible software frameworks and increased GPU computing helped make AI a practical tool for experimental and theoretical research.The passage situates this development alongside applications spanning protein structure prediction, weather forecasting, fusion-plasma control, and antibiotic discovery.Source: §1 · p. 2
  • AI and ML methods have been applied to condensed-matter wavefunctions, phase classification, materials-property prediction, and real-time experimental feedback.Examples include neural-network representations of quantum states, crystal-property prediction, and ML-guided scattering and microscopy experiments.Source: §1 · p. 2
  • Prior reviews cover ML in scattering and spectroscopy broadly, while an ARPES-focused review has concentrated mainly on CNN-based denoising and dispersion extraction.The present review extends coverage across the full ARPES workflow.Source: §1 · pp. 2–3
  • This review covers AI and ML across the complete ARPES workflow, from sample preparation and data acquisition to analysis and theoretical interpretation.It also examines the data landscape and prospects for autonomous experiments.Source: §1 · p. 3

2 Angle-Resolved Photoemission Spectroscopy

ARPES maps occupied electronic structure from photoelectrons, and modern measurements span many dimensions of momentum, energy, space, spin, time, and experimental conditions. AI-assisted processing, inversion, and autonomy address the growing complexity of acquisition and interpretation.

2 Angle-Resolved Photoemission Spectroscopy

  • Modern ARPES datasets span energy, momentum, spatial coordinates, spin, time, and controlled conditions, making manual real-time choices about what to measure increasingly time-consuming.Third-generation light sources with micro-focused spots under 50 µm enable spatial scans across domains within a sample.Source: §2 · p. 3
  • Photoelectron kinetic energy depends on photon energy, work function, and binding energy, while emission angle helps map crystal momentum.Changing photon energy provides access to perpendicular momentum.Source: §2 · p. 3
  • Within the sudden approximation, measured intensity is proportional to the spectral function multiplied by a photoemission matrix element.The matrix element depends on geometry, photon kinematics, and orbital character.Source: §2 · p. 4
  • The spectral function carries many-body physics through the Green’s function, and extracting bare-band dispersion and self-energy is a central ARPES inference goal.The real self-energy part renormalizes dispersion, while its imaginary part sets quasiparticle lifetime.Source: §2 · p. 4
  • A fixed-photon-energy ARPES rainbow plot is a curved 2D slice through (k, ω) space, so an energy-distribution curve need not correspond to a single momentum.Photoelectrons of a given kinetic energy lie on an emission sphere.Source: §2 · p. 4
  • AI opportunities fall into data processing, inversion of spectra into physical quantities, and autonomy for guiding acquisition and instrument control.Processing and inversion mainly support real-time or post-experiment analysis, while autonomy operates mainly during fabrication and experiments.Source: §2 · p. 4

3 Overview of ARPES workflow

The ARPES workflow comprises sample growth and preparation, measurement, and data analysis, supported by theory and complementary experimental probes. ML and AI connect these stages through automated measurement, rapid analysis, and data curation.

3 Overview of ARPES workflow

  • The workflow has three stages—sample growth and preparation, ARPES measurements, and data analysis—with theoretical modeling and complementary probes supporting material selection, experiment design, and interpretation.ML frameworks and AI agents are being incorporated across the workflow.Source: §3 · p. 4
  • Sample preparation includes identifying candidate materials and optimizing synthesis, with 2D vdW work extending to exfoliation, heterostructure assembly, and device fabrication.ARPES surface sensitivity requires rigorous control of cleanliness, contamination, and environment for air-sensitive materials.Source: §3 · p. 5
  • During measurement, preliminary data can classify spatial domains, assess quality, and guide where and how subsequent measurements are acquired [36] [37].Rapid processing such as domain labeling, feature identification, and peak fitting can feed results into the next acquisition cycle [36].Source: §3 · p. 5
  • Post-experiment analysis shifts from acquisition guidance to extracting electronic structure and interpreting material physics, while also supporting curation, metadata generation, and repository deposition.These data-management activities matter for reproducibility and future data-driven approaches.Source: §3 · p. 5

4 AI terminology for experimentalists

The review uses AI as a broad term for task-specific computational intelligence and distinguishes data-driven ML from explicitly programmed rules. It introduces model architectures and training protocols as separate ways to organize methods relevant to experimentalists.

4 AI terminology for experimentalists

  • In this review, AI refers broadly to task-specific artificial narrow intelligence rather than hypothetical artificial general intelligence.ML learns task-specific rules from data, whereas conventional symbolic AI relies on explicitly programmed instructions.Source: §4 · pp. 5–6
  • Traditional compact ML includes shallow statistical models such as PCA, NMF, t-SNE, and clustering, while deep learning uses multiple nonlinear layers to learn representations.Deep learning has improved image and object recognition and shows promise for spectroscopic maps.Source: §4 · p. 6
  • Learning protocols are commonly classified as supervised, unsupervised, self-supervised, or reinforcement learning, a separate axis from model architecture; in current practice, reinforcement learning is typically implemented with deep networks.Unsupervised methods infer structure without external labels, whereas supervised methods train on labeled data as ground truth.Source: §4 · p. 6, Figure 2 · p. 7
  • Self-supervised learning derives supervisory signals from unlabeled data, including through reconstruction, contrastive, or masked-image tasks.Examples include denoising and variational autoencoders, as well as DINO, SwAV, MoCo, and BYOL.Source: §4 · p. 7
  • Reinforcement learning trains an agent to maximize reward through interaction with an environment, without labeled targets.Such approaches are beginning to be applied to automated 2D-material fabrication [52].Source: §4 · p. 7
  • Physics-informed neural networks encode governing equations in the loss function, while multimodal models combine data types and generative models create new samples or representations.Generative-model predictions in scientific applications ultimately require validation against experimental measurements.Source: §4 · p. 7

5 Automation of sample preparation for ARPES measurements

Automation is advancing in sample synthesis and 2D-device fabrication, but producing ARPES-ready samples remains constrained by surface preparation and uncertain generalization beyond commonly demonstrated systems.

5 Automation of sample preparation for ARPES measurements

  • Bulk synthesis, thin-film deposition, surface preparation, heterostructure fabrication, contacting, and sample mounting are among the pre-measurement bottlenecks.Sample-preparation stages currently differ in their levels of automation.Source: §5 · pp. 7–8
  • The A-Lab combines recipe generation, robotic synthesis, X-ray diffraction feedback, and iterative optimization, producing dozens of inorganic powders on week-long timescales.Extending such workflows from powders to high-quality single crystals, epitaxial films, and well-ordered surfaces remains a challenge.Source: §5 · p. 8
  • Robotic 2D-material workflows automate steps from flake detection and stacking to closed-loop assembly, while pixel-by-pixel assembly controls layer number, composition, position, and twist angle.Approaches include computer-vision-guided pick-and-place, high-vacuum assembly, and platforms with closed-loop error correction.Source: §5 · pp. 8–9, Figure 3 · p. 8
  • Atomically clean, flat surfaces remain a critical rate-limiting step, and it is unresolved how well existing platforms generalize beyond graphene/hBN or how much expert intervention remains necessary.Related infrastructure efforts include the MonArk NSF Quantum Foundry, Argonne’s RAPID-200, and the Henry Royce Institute UHV 2D materials platform.Source: §5 · p. 9

6 ARPES experiment and its automation

AI-assisted ARPES acquisition spans sample handling, adaptive measurement, and spectral-quality assessment. These approaches can reduce acquisition time and guide measurement choices, but depend on meaningful learned representations and objectives.

Sample transport and positioning

  • Motorized transfer arms, manipulators, stages, and cleaving mechanisms reduce manual intervention in routine handling, providing hardware for future AI-guided procedures.At MAESTRO, computer-controlled transfer moves samples from the load lock toward cleaning/growth chambers and measurement chambers.Source: §6 · pp. 9–10

Adaptive acquisition

  • KMGP maps electronic domains from clustered spectra, while guided acquisition adaptively searches real and momentum space for regions with high spectral intensity and sharp features.The approaches use cluster labels or real-time communication with beamline control software to determine subsequent measurements.Source: §6 · p. 10, Figure 4 · p. 11
  • GPR reduced spin-ARPES acquisition time by 5–10 times versus empirical stopping criteria, using EDC-derived scores as a quantitative stopping rule.The reported application uses spin-polarization signals from Bi2Te3 spin-resolved EDCs.Source: §6 · p. 10

Spectral-quality assessment

  • A synthetic-spectrum-trained CNN can score spectral quality, identify misaligned flake domains, locate uniform regions, and determine relative orientations in twisted heterostructures.Its demonstration is limited to a material system, uses rule-based proxy labels, and randomizes rather than explicitly calculates matrix-element effects.Source: §6 · pp. 10–11

Spatial domain classification

  • Spatial domain classification can use pretrained image encoders with few-shot references or graph convolutional networks that promote neighboring spectra to share band dispersion.Direct k-means clustering of generic embeddings may perform poorly without carefully selected reference examples.Source: §6 · p. 11

Limitations and requirements

  • Adaptive sampling can miss important electronic features when cluster boundaries are ambiguous, learned representations are unfaithful, or optimization targets fail to reflect scientifically relevant information.Robust autonomy therefore requires expert supervision, physically meaningful constraints, uncertainty estimation, and adaptive decision-making within a specific objective and limited beamtime.Source: §6 · pp. 11–12

7 Post-experiment data analysis and insights

Post-experiment ARPES analysis applies image-based methods and quantitative inference to address low signal-to-noise ratios and overlapping bands. The reviewed approaches span denoising, spatial-domain mapping, band extraction, and inference of many-body parameters.

7.1 Denoising and spectral restoration

  • Liu et al. jointly disentangled intrinsic bands, Poisson noise, and collimator-grid artifacts, while denoising validation must account for simulation mismatch, over-smoothing, and false features.The review recommends validation against longer reference measurements or independent methods, and explicit labeling and documentation of processed spectra.Source: §7.1 · pp. 13–15
  • Cheng et al. achieved up to approximately 500-fold shorter acquisition; Kim et al. enabled analysis at two orders shorter dwell, while deep image prior yielded approximately 70-fold quantitative and up to 300-fold qualitative reductions.The reported reductions come from different methods and study conditions.Source: §7.1 · pp. 12, 14

7.2 Spatial-domain identification and clustering

  • Clustering maps spatial electronic domains, while multi-stage clustering and NMF-based activation clustering can expose inhomogeneity and mixed-phase contributions.Cluster parameters still require physical interpretation; weak contrasts, angular spread, or continuous boundaries can yield different partitions across algorithms or hyperparameters.Source: §7.2 · p. 15

7.3 Band Extraction and low-dimensional representations

  • For NbIrTe4, a CNN recovered band positions comparable to conventional extraction while reducing manual intervention and improving robustness on weak or overlapping features.The figure compares maximum-curvature tracing, minimum-gradient extraction, and CNN results against DFT and experimental data.Source: §7.3 · p. 16, Figure 7 · p. 16
  • An autoencoder compresses ARPES spectra into a reconstructable latent representation, while Bayesian MAP reconstruction recovered all 14 WSe2 valence bands using continuity priors.That prior can stabilize weak-signal reconstruction but may bias results toward the initial band structure; k-means also tracked differing carrier relaxation in PtBi2 states near Weyl points and the bulk Fermi surface.Source: §7.3 · pp. 16–17

7.4 From spectra to many-body parameters

  • Boltzmann-machine regression inferred normal and anomalous self-energies whose energy-dependent peaks largely cancel in the total spectral function; the anomalous peak accounted for more than 90% of the superconducting gap.The inversion requires high-SNR ARPES data and benefits from independent physical validation.Source: §7.4 · p. 17, Figure 8 · p. 18
  • Bayesian maximum-entropy xARPES addresses curved-dispersion self-energy extraction, while SIREN surrogates optimize tight-binding parameters against measured ARPES intensity.Traditional Lorentzian MDC fitting can introduce systematic bias for curved dispersions; SIREN uses automatic differentiation to optimize the parameters.Source: §7.4 · pp. 17–18

7.5 Transfer learning and future autonomous analysis

  • Transferable networks trained mainly on simulated spectra classified Bi2212 superconducting states from a single snapshot, and the review frames analysis as a modular toolkit rather than a universal ML solution.Future directions include standardized analysis pipelines and integration of spectroscopy-specific models with general-purpose reasoning tools and agentic systems.Source: §7.5 · pp. 18–19

8 Theory support: DFT surrogates and ARPES

ML surrogates can make theoretical electronic structures more available for ARPES comparison by shifting computation to training, but they inherit the accuracy limits of their reference calculations. Approaches range from Hamiltonian and quasiparticle models to long-range density extrapolation, with potential roles in analysis and acquisition.

8 Theory support: DFT surrogates and ARPES

  • Surrogate predictions inherit approximations and biases from their training calculations, so faster inference does not guarantee physically accurate quasiparticle bands.The review notes that a higher-fidelity teacher model may be needed to improve on a DFT-level surrogate's physical accuracy.Source: §8 · p. 19
  • Early density, DOS, and operator models speed electronic-structure estimates, but do not produce the momentum-resolved band maps required for ARPES.The section describes later DOS approaches across diverse chemistries, while identifying Hamiltonian or quasiparticle surrogates as necessary for E–k overlays.Source: §8 · pp. 19–20
  • DeepH learns a DFT Hamiltonian from atomic structure, bypasses SCF at inference, and reports meV-scale Hamiltonian accuracy for generating momentum-resolved bands.Its localized-orbital model exploits nearsightedness: matrix elements are limited to neighboring atoms within RC and depend on a local neighborhood within RN.Source: §8 · p. 20, Figure 9 · p. 21
  • Long-range density models trained on small bilayer cells extrapolate to large twisted moiré supercells and reproduce features such as flat bands, but the extrapolation needs validation.The stated validation concern is especially relevant when comparing with narrow, strongly twist-dependent ARPES bands.Source: §8 · p. 20
  • Knøsgaard and Thygesen predict self-energy corrections at roughly 0.14 eV MAE, while Hou et al. report about 0.11 eV MAE for GW quasiparticle structures.Both approaches transfer some higher-level quasiparticle calculation cost to trained surrogates; accuracy and transferability still depend on reference calculations.Source: §8 · p. 20
  • SlaKoNet learns Slater–Koster matrix elements from JARVIS-DFT data, retaining tight-binding interpretability while improving gaps relative to GGA and increasing GPU evaluation speed.It is a transparent DFT-level band model, not a GW surrogate.Source: §8 · pp. 20–21
  • Fast surrogates could support dispersion overlays and k⊥ interpretation after acquisition, and potentially guide photon-energy or polarization choices during acquisition.The review characterizes real-time theory-guided acquisition as less established than post-experiment applications; rapidly varying structures could also be tested without repeating full ab initio calculations.Source: §8 · p. 21

9 Data landscape: sporadic open repositories versus missing centralized ARPES database

ARPES data are available in open repositories, but remain fragmented and lack a large, shared, task-labeled resource. The review proposes a federated archive of pretrained task models and small benchmarks, while noting challenges in curation, ground truth, and adoption.

9 Data landscape: sporadic open repositories versus missing centralized ARPES database

  • An ARPES dataset needs task-specific labels and clear separation of raw, calibrated, processed, and labeled data because tasks use different inputs and targets.For example, denoising raw Poisson-limited counts differs from denoising second-derivative images.Source: §9 · p. 22
  • Public ARPES datasets remain fragmented, with large releases isolated rather than part of a universally maintained, multitask training resource.The review notes that public deposits often occur in response to institutional or funder requirements, while groups hold larger private archives.Source: §9 · p. 22
  • Useful repositories and FAIR metadata standards already exist, but their uneven adoption and missing common interfaces and data-streaming pathways constrain reuse and machine-learning integration.The review identifies beamline-specific control software and interfaces as an additional scaling bottleneck.Source: §9 · pp. 22–23
  • Longer acquisitions are not automatically ground truth because aging, drift, charging, and flux changes can make reference and short-acquisition measurements differ.The review recommends validating denoised or reconstructed spectra using physical quantities such as peak positions and linewidths, not visual smoothness alone.Source: §9 · p. 23
  • A federated open archive with small task-specific benchmarks and pretrained models is proposed as a practical route toward broader AI adoption alongside larger data deposits.Groups could fine-tune and redistribute models with cards describing task, architecture, training domain, failure modes, and update history.Source: §9 · p. 24
  • A practical starting point is to repackage existing task models and establish public test sets with agreed metrics; portability, versioning, stewardship, benchmarking, and training-data documentation also matter.The review warns that model cards without independent testing may be unreliable.Source: §9 · pp. 24–25

10 Perspective

The review frames ARPES automation as a progression in decision-making authority, while emphasizing that current systems remain mostly partial and that integrated autonomy faces technical and governance challenges. Its proposed goal keeps researchers focused on scientific judgment as routine workflow decisions become increasingly automated.

Towards the “Self-driving ARPES Lab”

Current Status and Future Trajectories

  • Most current ARPES systems fall between L1 and L3, and reviewed AI methods generally address individual tasks rather than control the complete experiment.Automation can vary across operations even within one endstation.Source: Current Status and Future Trajectories · p. 27
  • The inference barrier is deciding whether spectral features and data quality warrant a change in measurement conditions to gain useful information.Reliable decisions require better training data, standardized metadata, real-time data access, uncertainty estimates, and closer integration of beamline-control and analysis software.Source: Current Status and Future Trajectories · p. 27
  • Direct AI control of experimental hardware raises trust and control requirements, including defined operating limits, decision tracking, researcher understanding, and override.These safeguards matter particularly for expensive instrumentation and conditions that could damage a sample or device.Source: Current Status and Future Trajectories · p. 27
  • Fully integrated L4 or L5 ARPES experiments have not been demonstrated, although many component capabilities exist independently.The review identifies shared reusable models as a next step; larger organized datasets may help, while simulations, synthetic data, transfer learning, and continued learning during experiments are also routes, especially where experimental data are limited.Source: Current Status and Future Trajectories · p. 27
  • A self-driving ARPES lab is framed as real-time communication among sample handling, measurement, analysis, and theoretical calculations, not merely operation without a scientist.Routine decisions could shift to the automated system while researchers assess unexpected observations and whether the emerging picture makes physical sense.Source: Current Status and Future Trajectories · pp. 27–28

Competing interests

The paper reports no competing interests or newly generated datasets, describes author contributions, and explains how large language models were used in manuscript preparation.

Competing interests

Availability of data and material

Authors’ contributions

  • The listed authors contributed to project initiation and supervision, literature review, drafting, figures, scientific discussion, review, and revisions.Source: Authors’ contributions · pp. 28–29

Use of large language models

  • LLMs helped organize literature notes and occasionally improve clarity, grammar, and formatting under author direction; authors wrote the manuscript and retain responsibility for its scientific content.Source: Use of large language models · p. 29
Loading 2610.10140v1…