Strategy for Bayesian Optimized Beam Steering at TRIUMF-ISAC's MEBT and HEBT Beamlines

O. Hassan; O. Kester; O. Shelbaya; T. Planche; W. Fedorko

REVIEW 4 major objections 5 minor 23 references

Strategy for Bayesian Optimized Beam Steering at TRIUMF-ISAC's MEBT and HEBT Beamlines

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A black-box optimizer tunes post-accelerator beam steering section by section to 76–100% transmission in about 28 minutes.

desk verdict Honest, useful machine-development report; the headline numbers are best-of-runs composites, so the 'consistently' claim outruns the evidence. read the letter →

arxiv 2505.08767 v4 pith:4P6O36KP submitted 2025-05-13 physics.acc-ph physics.app-ph

O. Hassan , O. Shelbaya , W. Fedorko , T. Planche , O. Kester This is my paper · ORCID

classification physics.acc-phphysics.app-ph

keywords BayesianoptimizationbeamsteeringtransmissionacceleratortuningGaussianprocesssequencedecompositionrareisotopebeamscurrentmonitor

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper reports a strategy for automating beam steering in the medium- and high-energy transport lines of a rare-isotope post-accelerator. The authors claim that a Bayesian optimization routine, applied to overlapping beamline subsections, consistently delivers high transmission—between 76% and 100% across the five sequences tested—while keeping total tuning time near 28 minutes, comparable to manual tuning. The practical payoff is that accelerator operators could rely on a semi-automated, model-informed tuning procedure as the facility moves to simultaneous delivery of multiple rare-isotope beams.

What carries the argument

The workhorse is BOIS, a Bayesian optimizer that models the transmission objective as a Gaussian process with a smoothness kernel and chooses the next steerer settings through an acquisition function, either expected improvement or upper confidence bound. The key structural device is sequence decomposition: the beamline is split into overlapping sub-sections, each bounded by current monitors, so that each optimization has a small number of variables and a direct readback of the quantity to maximize. For the medium-energy-to-linac section, the optimized variables include quadrupole gradients bounded to within 10% of model values; for the high-energy sections only steerers are tuned. The paper's time estimate comes from manually stopping each sequence when transmission reaches 90–95% or stops improving.

What would settle it

Repeat the five-sequence tuning run while logging an upstream beam-current monitor continuously and compare the optimizer's reported transmission gains against source-current fluctuations: if per-sequence transmission tracks source current rather than steerer settings, the 28-minute estimate and the 76–100% figures do not hold. A cleaner test is to hold the ion source at fixed output and see whether Sequence 3 still reaches 93% transmission with the upper-confidence-bound acquisition function.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that beam steering through a post-accelerator's medium-energy, drift-tube-linac, and high-energy sections can be treated as a set of small black-box optimization problems, each with its own transmission objective read from a beam-current monitor. Dividing the machine into overlapping sequences keeps the number of steerer variables low (4 to 14), lets each subproblem use available diagnostics, and yields high transmission in every section. A single combined optimization with 14 steerers stalls at 68% transmission, whereas the sequential approach reaches 86% or better over the same stretch. The paper also finds that including quadrupoles in the optimization is necessary but only works when their gradients are bounded to within 10% of the model-computed values; unbounded quadrupoles collapse transmission to about 5%.

Load-bearing premise

The objective assumes the ratio of currents measured on two beam-current monitors is a stable signal that reflects steering quality; the paper records ion-source failures causing more than 20% current fluctuations and slow input-current drift, so if source drift is confounded with steerer response the reported gains could be partly an artifact of source conditions.

Editorial extensions

If this is right

Splitting a long beamline into overlapping diagnostic-bounded sequences lets a black-box optimizer beat a single big optimization: 86% transmission versus 68% over the same high-energy stretch.
With early stopping, full automated tuning from the medium-energy corner to the end of the high-energy section takes about 28 minutes, comparable to manual tuning time.
Bounding quadrupole gradients to within 10% of model-computed values is required for reliable linac injection; unbounded optimization reduces transmission to 5%.
The same sequence-based Bayesian strategy should transfer to any beamline with a continuous beam-current readback, including other injection lines and future superconducting-linac segments.
Strict version control of the optimizer software matters: an upgrade fixed a severe over-exploration failure in one sequence, changing transmission from 1–5% to 93%.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If source-current fluctuations are the main confound, then explicitly feeding the measured current-monitor variance into the Gaussian process, as the paper lists as future work, would make the 28-minute estimate and per-sequence transmissions more trustworthy.
The 90–95% early-stop threshold likely leaves a few percentage points of transmission on the table; the reported 76–100% range is therefore a floor under the method's performance, not its ceiling.
The same overlapping-subsequence recipe could be applied to other scalar beam objectives, such as emittance or energy spread, whenever a fast diagnostic readback exists.
A direct test of cross-species transfer would be seeding the initial sampling stage with best settings from previous runs scaled by mass-to-charge ratio, which the paper plans but has not yet demonstrated.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Quick take: this is a solid machine-development report from a real facility, not a new methods paper. The new content is empirical: transmission values for four species across five steering sequences, a sequence-splitting strategy for MEBT/HEBT, a clear demonstration that combining sequences into one 14-steerer optimization degrades performance (68% vs 86% for split sequences), and a documented BoTorch version failure. Those are useful, concrete results for anyone running Bayesian optimization on accelerator beamlines.

The paper earns credit for honesty. It reports the ion source failures, the >20% current fluctuations, the beam current drift from 6.2 to 5.5 nA, the manual early stopping, and the fact that sequence data were collected over several independent runs rather than one ordered sweep. That transparency is not typical and should be acknowledged.

The soft spots are real but mostly addressable. Table 2's caption says the data came from several independent runs, which means each row is a selected composite rather than an end-to-end measurement. The 28-minute total comes from a single 12C3+ test with manual early stopping; the paper itself says the pre-early-stopping time was about 42 minutes. There are no error bars, no repeat runs, and no controlled comparison against the operator baseline. So the summary sentence that BOIS 'consistently achieves high transmission rates' is stronger than the evidence supports. The source-drift concern is somewhat mitigated by the ratio-based Faraday cup objective and by testing during stable periods, but it is still a confounder for individual runs. The more direct gap is that no single continuous rerun demonstrates the full sequence with stopping criteria implemented. That is a fixable experimental gap, not a fundamental flaw.

The citation pattern looks fine. BOIS is the authors' prior tool from [10], and they cite it; the current paper extends it to new beamlines and species. No sign of inflated claims about novelty.

Who is this for? Accelerator physicists and controls people running Bayesian optimization at operational facilities. It would be useful reading before trying similar steering strategies elsewhere, and the quadrupole-bounding result (unbounded gives 5%, bounded gives 100% DTL transmission) is a practical warning that would be lost if this were desk-rejected.

My recommendation: send it to peer review. It deserves referee time, primarily because it reports credible measurements and honest limitations, but ask the authors to add at least one fully continuous end-to-end run with automatic early stopping, report uncertainties or variability across repeats, and soften the 'consistently' claim accordingly. With those changes this becomes a genuinely useful operational reference.

Referee Report

4 major / 5 minor

Summary. The manuscript reports machine development studies of Bayesian optimization (BOIS) applied to corrective steering for beam transmission at TRIUMF-ISAC's MEBT and HEBT lines. It argues that dividing the beamline into overlapping sub-sections (Sequences 1-5) and optimizing each with BOIS yields high transmission (76-100% per section in Table 2) and a total tuning time of about 28 minutes (Table 1) from the MEBT corner to HEBT2. The paper also documents failure cases (unbounded quadrupoles, over-exploration due to a BoTorch version bug) and outlines future steps for operational deployment.

Significance. If the reported performance is reproducible, the paper provides a practical, generalizable strategy for semi-automated beam steering at rare isotope facilities, with only a Faraday cup as diagnostic. The explicit sequence-splitting guidance and the honest reporting of failure modes are useful. However, the evidence as presented is not yet sufficient to support the 'consistently' claim, because the headline numbers are composites from several runs and lack uncertainties.

major comments (4)

[Section 4.6 / Table 2] The central claim that BOIS 'consistently' achieves high transmission rests on Table 2, but the table's own caption states the data were collected over several independent runs rather than a single, strictly ordered sweep. Each transmission value is therefore a best-of-runs result, and the reader cannot assess run-to-run variability. A single continuous end-to-end demonstration from Sequence 1 through Sequence 5, repeated at least a few times, is needed to support the 'consistently' claim and the 28-minute total.
[Section 4.4 / Table 1] The 28-minute tuning time was obtained with manual early stopping (stopping when transmission reached 90-95% or ceased to improve), while Section 5 lists automated early stoppage as future work. Hence the quoted time is not achievable with the current BOIS code as described, and no repeat runs or error bars are given. The paper should state clearly that this is a manual-stoppage estimate and provide a range or repeated trials.
[Sections 4.1 and 4.5] Ion source instability is a known confound: input current drifted from 6.2 to 5.5 nA during the MEBT tests (Section 4.1) and >20% current fluctuations occurred at irregular intervals in the later tests (Section 4.5). While the ratio-based Faraday cup objective partially mitigates common-mode source drift, the absence of quoted uncertainties on Table 2 entries and any statistical treatment of source fluctuations leaves the reported transmissions and the 28-minute time potentially influenced by source conditions. The authors should report uncertainties and state how stability periods were selected.
[Section 4.3] The evidence that sequence splitting is necessary rests on a single combined optimization (14 steerers) that reached only 68% transmission versus '86% achieved using smaller sequences.' This is a key quantitative comparison, but the 86% reference is not traceable to a particular row of Table 2, and no repeated runs or uncertainties are presented. Please clarify how the 86% was obtained and whether the comparison accounts for run-to-run variability.

minor comments (5)

[Table 1] Table 1 lists Sequence #1 with 14 elements while Sequences 2-5 have 4-6 elements; please define what counts as an element (steerers plus quadrupoles?) in the caption or text.
[Section 4.2] The statement that 'EI generally converges faster but is less consistent' is not quantitatively supported; consider adding a figure or a table with repeated runs to substantiate this claim.
[Section 2.1] The phrase 'inputs are randomly selected based on their mid-points' is ambiguous; please clarify whether the random sampling is uniform over the bounds or seeded from midpoint values.
[Figure 5 caption] In Figure 5, 'best transmission shown for each sequence' could mean either the best value during the run or the final value; please specify which is plotted.
[Section 4.6] The word 'consistently' in 'BOIS consistently achieves high transmission rates' is stronger than the evidence supports, as noted in the major comments; a more measured phrasing such as 'in the tests reported here' would be appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported transmissions and tuning times are measured against external beamline diagnostics, not derived from the optimizer's own assumptions or from a self-citation chain.

full rationale

The paper's central claim is that Bayesian Optimization for Ion Steering (BOIS), applied to suitably divided beamline sub-sections, achieves high measured transmission in short tuning times. The objective is explicitly external: Section 4.1 states 'The objective was to maximize beam transmission as measured from one Faraday cup to the next,' and the results in Table 2 are measured Faraday-cup transmissions for different species and sequences. There is no equation in which an output is an input by construction, and no fitted parameter is renamed as a prediction. The tuning-time estimate in Table 1 is a timed run with a stated manual early-stopping rule (Section 4.4: 'Manual stoppage was used when transmission reached 90-95% or ceased to improve'), so it is an empirical report rather than a derived forecast. The BOIS algorithm is attributed to the authors' prior work [10], but the present evidence is new machine data, not a result imported from that citation; the same-author references for the ±10% quadrupole bounds ([16,17]) and for the MEBT/HEBT models are contextual, and the paper's conclusions do not reduce to those citations. The main weaknesses are evidentiary rather than circular: Table 2's caption admits the sequence data came from 'several independent runs rather than in a single, strictly ordered sweep,' and Section 4.5 reports ion-source current fluctuations above 20%. These concerns bear on reproducibility and statistical strength, not on whether the derivation is equivalent to its inputs. Accordingly, no circularity step is identified.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central operational claim rests on three domain assumptions: transmission is the right objective, MCAT model values are trustworthy within chosen bounds, and ion-source fluctuations are separable from steering effects. The GP/BOIS machinery is standard. Hyperparameters beta and bounds are hand-chosen free parameters. No invented entities.

free parameters (4)

UCB acquisition function exploration coefficient beta = beta=3 (MEBT, HEBT seq 3), beta=10 (HEBT seq 2), EI beta=1 (seq 3 initial)
Chosen by hand from experiment trials; directly controls exploration versus exploitation and materially changes outcomes, including one failed over-exploration run.
Quadrupole gradient bounds = ±10% of MCAT-computed gradients
Choice of constraint from prior internal beam dynamics studies [16,17], not re-derived in this paper; MEBT optimization succeeds only with these bounds.
Early-stopping threshold = 90-95% transmission or no improvement
Manual stoppage used for timing estimates; directly affects the reported 28 minute total and success criterion.
Faraday cup averaging count = 20 measurements
Operational setting flagged as redundant in Future Work; affects the noise level of the objective function.

assumptions (4)

domain assumption Maximizing transmission measured on Faraday cups is a sufficient objective for tuning beamline steering
Used throughout; objective assumes transmission is the operational figure of merit and no other beam quality constraints apply.
domain assumption MCAT-computed quadrupole gradients are correct to within ±10%; deviations arise from model uncertainty
Needed for the MEBT sequence; previously investigated in internal reports [16,17], not independently validated in this paper.
standard math Gaussian process surrogate with Matérn kernel and chosen acquisition functions models the steering response well
BOIS relies on standard GP machinery from Rasmussen and Williams [13]; treated as standard.
domain assumption Beam current fluctuations during tests originate in the ion source and are separable from steering effects
Section 4.5 attributes >20% fluctuations to the ion source; if not separable, transmission gains could be confounded.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Strategy for Bayesian Optimized Beam Steering at TRIUMF-ISAC's MEBT and HEBT Beamlines." pith.science (2026). https://pith.science/paper/4P6O36KP

@misc{pith2026250508767,
  author       = {Pith},
  title        = {Pith review of: Strategy for Bayesian Optimized Beam Steering at TRIUMF-ISAC's MEBT and HEBT Beamlines},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4P6O36KP}},
  note         = {Machine review of arXiv:2505.08767}
}

read the original abstract

In preparation for operation of multiple Rare Isotope Beams (RIBs) when the Advanced Rare Isotope Laboratory (ARIEL) becomes operational, TRIUMF embarked on a program of advanced beam tuning applications and machine learning tools. The strategy for operationalizing Bayesian Optimization for beam steering purposes is being developed. A previously reported centroid correction algorithm is used to tune accelerated charged particle beams at TRIUMF's ISAC postaccelerator facility. We present findings and results from multiple machine development experiments conducted between October and November 2024, as part of a pivot toward semi-automated machine tuning methods. These findings were instrumental in shaping the tuning strategy for the medium and high energy beam transport (MEBT, HEBT) lines at ISAC, by sequentially optimizing sub-sections of the beamlines.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 21 canonical work pages

[10]

Ghelfi, A

E. Ghelfi, A. Katrusiak, R. Baartman, W. Fedorko, O. Kester, G. Kogler Anele et al.,Bayesian optimization for ion beam centroid correction,Review of Scientific Instruments96(2025) 023304

work page 2025
[1]

G.C. Ball, G. Hackman and R. Krücken,The triumf-isac facility: two decades of discovery with rare isotope beams, Physica Scripta91(2016) 093002

work page 2016
[2]

Koscielniak, R

S. Koscielniak, R. Laxdal, R. Lee and L. Root,Beam dynamics studies on the ISAC RFQ at TRIUMF, in Proceedings of the 1997 Particle Accelerator Conference (Cat. No.97CH36167), vol. 1, pp. 1102–1104 vol.1, 1997, DOI

work page 1997
[3]

Laxdal, G

R. Laxdal, G. Dutto, K. Fong, G. Mackenzie, M. Pasini, R. Poirier et al.,Beam commissioning and first operation of the ISAC DTL at TRIUMF, inPACS2001. Proceedings of the 2001 Particle Accelerator Conference (Cat. No.01CH37268), vol. 5, pp. 3942–3944 vol.5, 2001, DOI

work page 2001
[4]

Jayamanna,Off line ion source terminal, inISAC and ARIEL: The TRIUMF Radioactive Beam Facilities and the Scientific Program, J

K. Jayamanna,Off line ion source terminal, inISAC and ARIEL: The TRIUMF Radioactive Beam Facilities and the Scientific Program, J. Dilling, R. Krücken and L. Merminga, eds., (Dordrecht), pp. 51–62, Springer Netherlands (2014), DOI

work page 2014
[5]

Laxdal, P

R. Laxdal, P. Bricault, T. Reis and D. Gorelov,A separated function drift-tube linac for the isac project at triumf, inProceedings of the 1997 Particle Accelerator Conference (Cat. No. 97CH36167), vol. 1, pp. 1194–1196, IEEE, 1997

work page 1997
[6]

Dilling, R

J. Dilling, R. Krücken and L. Merminga,Ariel overview, inISAC and ARIEL: The TRIUMF Radioactive Beam Facilities and the Scientific Program, J. Dilling, R. Krücken and L. Merminga, eds., (Dordrecht), pp. 253–262, Springer Netherlands (2014), DOI. – 9 –

work page 2014
[7]

Shelbaya, R

O. Shelbaya, R. Baartman and O. Kester,Fast radio frequency quadrupole envelope computation for model based beam tuning, Phys. Rev. Accel. Beams22(2019) 114602

work page 2019

Show all 23 references

[8]

Shelbaya, T

O. Shelbaya, T. Angus, R. Baartman, P.M. Jung, O. Kester, S. Kiy et al.,Autofocusing drift tube linac envelopes,Phys. Rev. Accel. Beams24(2021) 124602

2021
[9]

Shelbaya, R

O. Shelbaya, R. Baartman, P. Braun, P. Jung, O. Kester, T. Planche et al.,Tuning methods for multigap drift tube linacs,Review of Scientific Instruments95(2024)

2024
[11]

Baartman,TRANSOPTR: Changes since 1984, Tech

R. Baartman,TRANSOPTR: Changes since 1984, Tech. Rep. TRI-BN-16-06, TRIUMF (2016)

2016
[12]

Shelbaya,Model Coupled Accelerator Tuning Software, Tech

O. Shelbaya,Model Coupled Accelerator Tuning Software, Tech. Rep. TRI-DN-25-01, TRIUMF (January, 2025)

2025
[13]

Rasmussen and C.K.I

C.E. Rasmussen and C.K.I. Williams,Gaussian Processes for Machine Learning (Adaptive Computation and Machine Learning), The MIT Press (2005)

2005
[14]

Matérn,Spatial Variation, Springer-Verlag, 2 ed

B. Matérn,Spatial Variation, Springer-Verlag, 2 ed. (1960)

1960
[15]

Frazier,A tutorial on bayesian optimization,ArXiv abs/1807.02811 (2018)

P. Frazier,A tutorial on bayesian optimization,ArXiv abs/1807.02811 (2018)

2018 arXiv
[16]

Shelbaya,Beam Dynamics Study of ISAC-MEBT, Tech

O. Shelbaya,Beam Dynamics Study of ISAC-MEBT, Tech. Rep. TRI-BN-22-29, TRIUMF (2022)

2022
[17]

Shelbaya,Model Coupled Accelerator Tuning, Ph.D

O. Shelbaya,Model Coupled Accelerator Tuning, Ph.D. thesis, The University of Victoria, Dept. of Physics and Astronomy, 2, 2023

2023
[18]

Srinivas, A

N. Srinivas, A. Krause, S. Kakade and M. Seeger,Gaussian process optimization in the bandit setting: No regret and experimental design, pp. 1015–1022, 07, 2010

2010
[19]

Hassan,Polarity Switching at the HEBT Steerers, Tech

O. Hassan,Polarity Switching at the HEBT Steerers, Tech. Rep. TRI-BN-24-12, TRIUMF (2024)

2024
[20]

Mockus, V

J. Mockus, V. Tiesis and A. Zilinskas,The application of bayesian methods for seeking the extremum, vol. 2, pp. 117–129, North-Holland (2014)

2014
[21]

Shelbaya,TRANSOPTR Implementation of the HEBT Beamlines, Tech

O. Shelbaya,TRANSOPTR Implementation of the HEBT Beamlines, Tech. Rep. TRI-BN-19-06, TRIUMF (2019)

2019
[22]

Balandat, B

M. Balandat, B. Karrer, D.R. Jiang, S. Daulton, B. Letham, A.G. Wilson et al.,Botorch: A framework for efficient monte-carlo bayesian optimization, 2020

2020
[23]

C. Charles et al.,The astrochemical and planetary materials irradiation experiment (ASPIRE) at TRIUMF, inProceedings of the 56th Lunar and Planetary Science Conference (LPSC 2025), (Woodlands, TX), 2025. – 10 –

2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [10]

Ghelfi, A

E. Ghelfi, A. Katrusiak, R. Baartman, W. Fedorko, O. Kester, G. Kogler Anele et al.,Bayesian optimization for ion beam centroid correction,Review of Scientific Instruments96(2025) 023304

work page 2025

[2] [1]

G.C. Ball, G. Hackman and R. Krücken,The triumf-isac facility: two decades of discovery with rare isotope beams, Physica Scripta91(2016) 093002

work page 2016

[3] [2]

Koscielniak, R

S. Koscielniak, R. Laxdal, R. Lee and L. Root,Beam dynamics studies on the ISAC RFQ at TRIUMF, in Proceedings of the 1997 Particle Accelerator Conference (Cat. No.97CH36167), vol. 1, pp. 1102–1104 vol.1, 1997, DOI

work page 1997

[4] [3]

Laxdal, G

R. Laxdal, G. Dutto, K. Fong, G. Mackenzie, M. Pasini, R. Poirier et al.,Beam commissioning and first operation of the ISAC DTL at TRIUMF, inPACS2001. Proceedings of the 2001 Particle Accelerator Conference (Cat. No.01CH37268), vol. 5, pp. 3942–3944 vol.5, 2001, DOI

work page 2001

[5] [4]

Jayamanna,Off line ion source terminal, inISAC and ARIEL: The TRIUMF Radioactive Beam Facilities and the Scientific Program, J

K. Jayamanna,Off line ion source terminal, inISAC and ARIEL: The TRIUMF Radioactive Beam Facilities and the Scientific Program, J. Dilling, R. Krücken and L. Merminga, eds., (Dordrecht), pp. 51–62, Springer Netherlands (2014), DOI

work page 2014

[6] [5]

Laxdal, P

R. Laxdal, P. Bricault, T. Reis and D. Gorelov,A separated function drift-tube linac for the isac project at triumf, inProceedings of the 1997 Particle Accelerator Conference (Cat. No. 97CH36167), vol. 1, pp. 1194–1196, IEEE, 1997

work page 1997

[7] [6]

Dilling, R

J. Dilling, R. Krücken and L. Merminga,Ariel overview, inISAC and ARIEL: The TRIUMF Radioactive Beam Facilities and the Scientific Program, J. Dilling, R. Krücken and L. Merminga, eds., (Dordrecht), pp. 253–262, Springer Netherlands (2014), DOI. – 9 –

work page 2014

[8] [7]

Shelbaya, R

O. Shelbaya, R. Baartman and O. Kester,Fast radio frequency quadrupole envelope computation for model based beam tuning, Phys. Rev. Accel. Beams22(2019) 114602

work page 2019