arxiv: 2604.05914 · v1 · submitted 2026-04-07 · ⚛️ physics.acc-ph

Recognition: no theorem link

Machine-State Embeddings as an Operational Coordinate System for Accelerator Operation

Chris Tennant, Jundong Li, Song Wang

Pith reviewed 2026-05-10 18:09 UTC · model grok-4.3

classification ⚛️ physics.acc-ph

keywords graph neural networksaccelerator operationmachine state embeddingsCEBAF injectordensity-based clusteringoperational monitoringanomaly detectionregime persistence

0 comments

The pith

Graph neural network embeddings of CEBAF injector snapshots create a 16-dimensional space where operations cluster into ten stable, persistent neighborhoods rather than a featureless continuum.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that GNN embeddings turn raw injector configuration data into an operational coordinate system for the CEBAF accelerator. Across 137,389 snapshots collected over more than a year, the embeddings reveal that the machine occupies only a handful of well-separated regions that correspond to actual operating regimes. These regions persist for hours to weeks, large jumps between them are rare, and the structure supports outlier detection and historical case retrieval. A sympathetic reader would care because single-sensor monitoring misses the holistic state, while this embedding approach gives a compact, interpretable map that aligns with real operations and can flag when the injector leaves normal behavior.

Core claim

We demonstrate that graph neural network embeddings of injector configurations provide a practical operational coordinate system for the CEBAF injector at Jefferson Lab. Using 137,389 snapshots spanning January 2022 through March 2023, we show that injector operation occupies a small number of persistent, well-separated neighborhoods in a 16-dimensional learned state space rather than a featureless continuum. Density-based clustering identifies ten recurring operating regimes with strong operational run alignment, and regime persistence statistics confirm that these regimes are stable over timescales of hours to weeks. Large relocations between neighborhoods are rare and episodic; 99.6% of 1

What carries the argument

The 16-dimensional GNN embedding space derived from injector configuration snapshots, which functions as an operational coordinate system by exposing clustered neighborhoods that correspond to distinct, stable operating regimes.

If this is right

Density-based clustering on the embeddings yields ten recurring operating regimes that align strongly with actual run records.
Regime persistence statistics show that 99.6% of one-hour windows remain inside an empirically derived jitter baseline.
Geometric outlier screening on the embedding space narrows a year-long archive to a small set of intervals that warrant operational review.
Nearest-neighbor lookup in embedding space enables case-based reasoning by retrieving similar past configurations from the historical record.
A controlled beam study confirms that deliberate injector reconfigurations produce coherent, interpretable trajectories through the embedding space.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same embedding technique could be applied to other accelerator sections or entire machines to create unified state maps across subsystems.
Embedding-space distances might serve as a quantitative metric for how far a current configuration has drifted from established good regimes.
The approach provides a natural way to test whether new operating procedures stay within previously observed neighborhoods or explore genuinely new states.
Connecting to high-dimensional sensor anomaly detection in other complex systems such as fusion devices or manufacturing lines.

Load-bearing premise

The GNN embeddings learned from the 137k snapshots capture operationally meaningful and stable states rather than artifacts of data collection, sensor noise, or model training choices.

What would settle it

Embed a fresh collection of injector snapshots with the same trained GNN model and check whether the points form a featureless continuum without ten persistent, run-aligned clusters or whether regime persistence drops below the reported hourly-to-weekly stability.

Figures

Figures reproduced from arXiv: 2604.05914 by Chris Tennant, Jundong Li, Song Wang.

**Figure 2.** Figure 2: PCA projection colored by point density (left) and beam current (right). Dense cores mark frequently occupied [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: HDBSCAN clusters identified in the 16D embedding space and visualized in the 2D PCA projection. Cluster [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Time distribution of cluster membership shown as a stacked histogram. Long contiguous blocks indicate that [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Distribution of step sizes for the 884 shift-to-shift transitions, computed as the Euclidean distance between [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Shift-to-shift embedding-space step sizes as a function of time. The dashed line marks the 95th-percentile [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

**Figure 7.** Figure 7: Distance of each shift centroid from the final shift (March 20, 2023), plotted over time. Breaks indicate [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗

**Figure 8.** Figure 8: Stability ratio (reading-space dispersion normalized to the pooled anchor-window [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗

**Figure 9.** Figure 9: Stability ratio versus mean beam current per one-hour window. Elevated ratios occur primarily at very low [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗

**Figure 10.** Figure 10: Outliers (red) overlaid on the PCA projection. Flagged points lie preferentially at regime boundaries and [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗

**Figure 11.** Figure 11: Historical-neighbor retrieval for three query types: a point from a large persistent cluster (upper left), a point [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗

**Figure 12.** Figure 12: PCA projection of the 62-step controlled beam study. Start and end configurations are marked by red [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗

read the original abstract

We demonstrate that graph neural network (GNN) embeddings of injector configurations provide a practical operational coordinate system for the Continuous Electron Beam Accelerator Facility (CEBAF) injector at Jefferson Lab. Using 137,389 snapshots spanning January 2022 through March 2023, we show that injector operation occupies a small number of persistent, well-separated neighborhoods in a 16-dimensional learned state space rather than a featureless continuum. Density-based clustering identifies ten recurring operating regimes with strong operational run alignment, and regime persistence statistics confirm that these regimes are stable over timescales of hours to weeks. Large relocations between neighborhoods are rare and episodic; 99.6% of one-hour operating windows fall within an empirically derived jitter baseline. Geometric outlier screening narrows a year-long dataset to a small set of intervals warranting operational review, and nearest-neighbor retrieval enables case-based reasoning over the historical archive. A controlled beam study validates that deliberate injector reconfiguration traces coherent, interpretable trajectories in embedding space. Together these capabilities demonstrate that machine-state embeddings support holistic operational monitoring in ways that single-channel inspection cannot.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

GNN embeddings turn CEBAF injector snapshots into ten stable operational regimes with run alignment, but the work skips the checks needed to show the separation is not an artifact of the model.

read the letter

The main thing to know is that this paper shows injector operation at CEBAF sits in a handful of persistent neighborhoods in a learned 16D space rather than filling a continuum. Using 137k snapshots they cluster into ten regimes that line up with actual runs and stay put over hours to weeks, with a controlled beam test tracing coherent paths through the space. That gives operators a practical way to monitor the whole machine state at once instead of single channels, plus outlier flagging and nearest-neighbor lookup for past cases.

Referee Report

2 major / 2 minor

Summary. The manuscript demonstrates that graph neural network embeddings of CEBAF injector configurations, derived from 137,389 snapshots over 14 months, form a 16-dimensional operational coordinate system. Density-based clustering identifies ten persistent, well-separated regimes with strong run alignment and stability over hours-to-weeks timescales; 99.6% of one-hour windows remain within a jitter baseline, large relocations are rare, and a controlled beam study shows coherent trajectories. The approach enables outlier screening, nearest-neighbor retrieval for case-based reasoning, and holistic monitoring beyond single-channel inspection.

Significance. If the embeddings prove robust to architecture and data choices, the work supplies a practical, data-driven state space for accelerator operation that supports regime identification, persistence tracking, and anomaly flagging at scale. The combination of large historical dataset, controlled validation, and operational utilities (outlier narrowing, case retrieval) distinguishes it from purely descriptive clustering studies and could inform monitoring tools at similar facilities.

major comments (2)

[Abstract and §4] Abstract and §4 (clustering results): the claim that injector operation occupies 'well-separated neighborhoods rather than a featureless continuum' rests on density-based clustering yielding exactly ten regimes, yet no quantitative validation (silhouette score, Davies-Bouldin index, inter-cluster separation statistics, or stability under bootstrap resampling) is reported. This metric gap is load-bearing because the skeptic concern—that GNN message-passing and 16D embedding choices may induce apparent discreteness—cannot be assessed without these numbers.
[§5] §5 (controlled beam study): trajectories are described as 'coherent and interpretable,' but the section supplies no error bars, quantitative alignment metrics between observed embedding paths and expected parameter changes, or ablation against noise-augmented or randomized snapshot subsets. Without these, the study does not directly test whether the observed structure survives perturbations that would be expected if the geometry were artifactual.

minor comments (2)

[Abstract] The abstract states '99.6% of one-hour operating windows fall within an empirically derived jitter baseline' but does not define the baseline computation or the precise windowing procedure; a short methods paragraph or equation would clarify reproducibility.
[§2] Dataset collection details (operator bias in snapshot timing, sensor calibration drift over the 14-month span) are mentioned only in passing; a dedicated subsection on data provenance would strengthen the claim that the 137k snapshots represent unbiased operational coverage.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We agree that the manuscript would be strengthened by the addition of quantitative validation metrics for both the clustering analysis and the controlled beam study. We have revised the manuscript to incorporate these elements and address the concerns directly. Below we respond to each major comment.

read point-by-point responses

Referee: [Abstract and §4] Abstract and §4 (clustering results): the claim that injector operation occupies 'well-separated neighborhoods rather than a featureless continuum' rests on density-based clustering yielding exactly ten regimes, yet no quantitative validation (silhouette score, Davies-Bouldin index, inter-cluster separation statistics, or stability under bootstrap resampling) is reported. This metric gap is load-bearing because the skeptic concern—that GNN message-passing and 16D embedding choices may induce apparent discreteness—cannot be assessed without these numbers.

Authors: We agree that the absence of these quantitative metrics leaves the discreteness claim vulnerable to the interpretation that it could be an artifact of the GNN message-passing or the 16-dimensional embedding. In the revised manuscript we have added a new subsection to §4 that reports the silhouette score, Davies-Bouldin index, inter-cluster separation statistics, and results from bootstrap resampling (100 iterations at 80% subsampling). These additions demonstrate that the ten regimes remain stable and well-separated, directly addressing the concern that the observed structure may be induced by modeling choices rather than reflecting genuine operational regimes. revision: yes
Referee: [§5] §5 (controlled beam study): trajectories are described as 'coherent and interpretable,' but the section supplies no error bars, quantitative alignment metrics between observed embedding paths and expected parameter changes, or ablation against noise-augmented or randomized snapshot subsets. Without these, the study does not directly test whether the observed structure survives perturbations that would be expected if the geometry were artifactual.

Authors: We acknowledge that §5 would be more convincing with error bars, quantitative alignment metrics, and ablation studies to test robustness against perturbations. The revised §5 now includes error bars on the trajectory plots (derived from repeated controlled runs), a quantitative alignment metric correlating embedding displacements with the magnitude of deliberate parameter changes, and an ablation comparing the observed trajectories against noise-augmented and fully randomized snapshot subsets. The results show that coherent trajectories persist under realistic noise but degrade under randomization, providing direct evidence that the geometry is not artifactual. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical discovery of neighborhoods in learned embedding space

full rationale

The paper trains a GNN on 137k historical snapshots to produce 16-dimensional embeddings, then applies density-based clustering and persistence statistics to identify ten operating regimes. All central claims (neighborhood separation, run alignment, 99.6% jitter compliance, trajectory coherence in the controlled study) are direct empirical observations on the output of this pipeline rather than quantities defined in terms of themselves or fitted parameters renamed as predictions. No self-citation chains, ansatzes, or uniqueness theorems are invoked to justify the existence or stability of the regimes; the geometry is measured from the data. The derivation chain therefore remains non-circular and externally falsifiable against the raw snapshot archive.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard machine-learning assumptions about data representativeness and embedding quality plus domain assumptions about injector operation stability.

free parameters (1)

GNN hyperparameters and embedding dimension
16-dimensional space and network architecture choices are fitted or selected during training on the 137k snapshots.

axioms (1)

domain assumption The collected snapshots are representative of normal and recurring injector operation
Invoked implicitly when claiming persistent regimes from the January 2022–March 2023 data.

pith-pipeline@v0.9.0 · 5488 in / 1214 out tokens · 38948 ms · 2026-05-10T18:09:13.861642+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 12 canonical work pages

[1]

doi:10.1109/MIM.2021.9620023. M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander. LOF: Identifying density-based local outliers. InProc. ACM SIGMOD Int. Conf. on Management of Data, volume 29, pages 93–104,

work page doi:10.1109/mim.2021.9620023 2021
[2]

doi:10.1145/335191.335388. R. J. Campello, D. Moulavi, A. Zimek, and J. Sander. Hierarchical density estimates for data clustering, visualization, and outlier detection.ACM Transactions on Knowledge Discovery from Data (TKDD), 10(1):1–51,

work page doi:10.1145/335191.335388
[3]

10 Jacob Cohen

doi:10.1007/978-3-642-37456-2_14. 16 Machine-State Embeddings as an Operational Coordinate System for Accelerator Operation T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. InProc. 37th Int. Conf. on Machine Learning (ICML), pages 1597–1607,

work page doi:10.1007/978-3-642-37456-2_14
[4]

doi:10.1146/annurev-nucl-121423-100719

ISSN 1545-4134. doi:10.1146/annurev-nucl-121423-100719. URL http://dx.doi.org/10.1146/ annurev-nucl-121423-100719. E. Fol, R. Tomás, J. Coello de Portugal, and G. Franchetti. Detection of faulty beam position monitors using unsupervised learning.Phys. Rev. Accel. Beams, 23:102805,

work page doi:10.1146/annurev-nucl-121423-100719
[5]

doi:10.1103/PhysRevAccelBeams.23.102805. Y . Gorishniy, I. Rubachev, V . Khrulkov, and A. Babenko. Revisiting deep learning models for tabular data. InAdvances in Neural Information Processing Systems (NeurIPS), volume 34, pages 18932–18943,

work page doi:10.1103/physrevaccelbeams.23.102805
[6]

doi:10.1109/TBDATA.2019.2921572. I. T. Jolliffe.Principal Component Analysis. Springer, New York, NY , USA, 2nd edition,

work page doi:10.1109/tbdata.2019.2921572 2019
[7]

doi:10.1016/j.conengprac.2019.104189

ISSN 0967-0661. doi:10.1016/j.conengprac.2019.104189. URLhttps://www.sciencedirect.com/science/article/pii/S0967066119301728. D. P. Kingma and M. Welling. Auto-Encoding variational Bayes. InProc. 2nd Int. Conf. on Learning Representations (ICLR),

work page doi:10.1016/j.conengprac.2019.104189 2019
[8]

doi:10.1109/ICDM.2008.17. L. McInnes, J. Healy, N. Saul, and L. Großberger. UMAP: Uniform manifold approximation and projection for dimension reduction.J. Open Source Softw., 3(29):861,

work page doi:10.1109/icdm.2008.17 2008
[9]

doi:10.21105/joss.00861. S. Ramaswamy, R. Rastogi, and K. Shim. Efficient algorithms for mining outliers from large data sets. InProc. ACM SIGMOD Int. Conf. on Management of Data, volume 29, pages 427–438,

work page doi:10.21105/joss.00861
[10]

doi:10.1145/335191.335437. L. van der Maaten and G. Hinton. Visualizing data using t-SNE.J. Mach. Learn. Res., 9(86):2579–2605,

work page doi:10.1145/335191.335437
[11]

doi:10.3389/fdata.2024.1366469. P. Wójcik and T. Barszcz. Condition monitoring using a latent space of variational autoencoder trained on healthy data. Sensors, 24(21):6670,

work page doi:10.3389/fdata.2024.1366469 2024
[12]

doi:10.3390/s24216825. Y . You, T. Chen, Y . Sui, T. Chen, Z. Wang, and Y . Shen. Graph contrastive learning with augmentations. InAdvances in Neural Information Processing Systems (NeurIPS), volume 33, pages 5812–5823,

work page doi:10.3390/s24216825
[13]

12 Supplementary Methods: Logbook Analysis Workflow For the September 30, 2022 case study, logbook entries were analyzed to connect a large embedding-space relocation event to operator-recorded machine context. The goal of this workflow was not to automate root-cause attribution, but to assemble a chronologically ordered, human-authored narrative record t...

2022