id stringlengths 36 36 | text stringlengths 1 1k | source stringlengths 42 191 | page int64 0 1.03k |
|---|---|---|---|
49150e7f-2aad-4cba-84cc-d7d2f92920f6 | voltageonaper-turnbasis . Amorerealisticapproachisto
adjust the generator current phasor ˜𝐼𝑔via a PI controller, as
commonly implemented in low-level RF feedback systems.
For implementation in a macro-particle tracking code,
thecavityvoltagephasorisupdatedbucketbybucket(i.e.,
every RF period) using Eq. (4). The remai... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 1 |
9ef6e5c1-213e-4346-bf09-e1b3a24601bf | directlytotherealandimaginaryparts. Toaccuratelymodel
LLRF feedback for beam dynamics simulations, it is essen-
tialto follow arealistic signalprocessing chain. This work
references the LLRF system at the SSRF storage ring [6],
as shown in Fig. 2. This system comprises four control
loops: amplitude,phase,tuning,anddire... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 1 |
a35cd478-ba65-476f-89d7-2ce0558dfe7c | eral stages. The cavity voltage signal (e.g., 500 MHz) is
picked up, down-converted to an intermediate frequency
(IF, e.g., 31.25 MHz), and then sampled by an Analog-to-
DigitalConverter(ADC)ataclockfrequency(e.g.,25MHz).
The ADC samples are transformed into I/Q components us-
ing a non-I/Q demodulation algorithm (with... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 1 |
beb0ba7b-308e-497b-85d3-408a3143dba5 | values. Finally, these corrected I/Q signals are modulated
backtoanIFdigitalsignal,convertedtoananalogsignalby
aDigital-to-AnalogConverter(DAC),up-convertedtothe
RF frequency, amplified, and used to drive the cavity.17th International Particle Accelerator Conference, Deauville, France JACoW Publishing
ISBN: 978-3-95450... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 1 |
44ee0832-ff6d-4bc4-9c1f-44da5cf86dfe | Basedonthisworkflow,weproposeasimplifiedyetrealis-
ticmodelforthefieldloop,depictedinFig.3. Thefeedback
procedure in our simulation is divided into three sequential
stages.
Figure3: SchematicofthemodeledLLRFcontrolsystem
for simulation.
Stage 1: Before the PI Controller
The 64-tap FIR filter operation is equivalent to ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 2 |
df282212-8dbc-4cd2-862e-bcddde253e61 | makingitstraightforwardtocomputeamovingaverageover
the last N buckets. In practice, this parameter N can be
estimated based on the specific LLRF system configuration
anddirectlysetwithintheSTABLEcodeaccordingly. The
voltage error phasor Δ˜𝑉is generated by subtracting this
averaged voltage phasor from the reference set... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 2 |
db6caa4d-3da5-4888-b70f-c3654ebf6e75 | gralgain,respectively. Inthediscrete-timesimulation,the
PI controller is updated every 𝑇𝑠and the integral term is
implemented by a forward-Euler accumulator:
𝐼𝑛=𝐼𝑛−1+𝐾𝑖,cont𝑇𝑠𝑒𝑛, (5)
and
(Δ˜𝐼𝑔)𝑛=𝐾𝑝𝑒𝑛+𝐼𝑛. (6)Equivalently, by defining the discrete-time integral gain
𝐾𝐼≜𝐾𝑖,cont𝑇𝑠, the above equa... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 2 |
c7dab34b-874b-449a-9ea3-f6ec602f1bf9 | conditions:
˜𝐼𝑔0𝑅𝐿cos(𝜓)𝑒−𝑗𝜓=˜𝑉𝑔0=˜𝑉set−⟨˜𝑉𝑏0⟩,(8)
where𝜓is the cavity detuning angle, ˜𝑉setis the desired
cavityvoltage,and⟨˜𝑉𝑏0⟩istheinitialaveragebeam-induced
voltage phasor across all buckets.
Thetotalgeneratorcurrentphasor, ˜𝐼𝑔=Δ˜𝐼𝑔+˜𝐼𝑔0,isthen
subjected to a delay before driving the cavity ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 2 |
9f1dff63-6c72-499f-bf3f-8f16963661b6 | sources, wheresignificant bunchlengthening (morethana
factor of 4) via harmonic cavities is desired. This lengthen-
ing can potentially introduce instabilities that challenge the
performance of the main RF cavity’s DLLRF system.
This model has been successfully applied to simulate and
analyze how the main RF cavity’s P... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 2 |
1bc673de-0d5b-4063-96d5-ae15778b14a9 | scope of the present paper and is planned for future work.
ACKNOWLEDGEMENTS
ThisworkwassupportedbytheFundamentalResearch
Funds for the Central Universities (No. WK2310000127)
and the National Natural Science Foundation of China
(Grants No. 12375324 and No. 12105284).17th International Particle Accelerator Conference, D... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 2 |
1821422a-88a4-4703-bab4-833c88debd60 | REFERENCES
[1]T. G. Berenc and M. Borland, “Modeling RF Feedback in Ele-
gant for Bunch-Lengthening Studies for the Advanced Photon
Source Upgrade”, in Proc. IPAC’15 , Richmond, VA, USA,
May 2015, pp. 540–542.
doi:10.18429/JACoW-IPAC2015-MOPMA006
[2]S.Wang et al.,“RFFeedbackSimulationforDiamond-IIUs-
ingELEGANT”,in Pro... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 3 |
96424beb-27df-4742-b2c5-7ca706fa9994 | doi:10.1103/PhysRevAccelBeams.21.012001
[5]T.HeandZ.Bai,“Graphics-Processing-Unit-AcceleratedSim-
ulation for Longitudinal Beam Dynamics of Arbitrary Bunch
Trains in Electron Storage Rings”, Phys. Rev. Accel. Beams ,
vol. 24, no. 10, p. 104401, Oct. 2021.
doi:10.1103/PhysRevAccelBeams.24.104401
[6]Y.-Y. Xia, Z.-T. Zhao... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 3 |
2f5c8238-84c8-41d8-8b65-8fb45621861f | ISBN: 978-3-95450-252-3 ISSN: 2673-5350 doi: 10.18429/JACoW-IPAC2026-WEP5602
MC5.D05: Coherent and Incoherent Instabilities Theory, Simulations, Code Development 2603WEP: Poster session WEP5602
Content from this work may be used under the terms of the CC BY 4.0 licence (© 2026). Any distribution of this work must maint... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP5602.pdf | 3 |
83f1e1a2-d53d-48c4-a30f-413febb50a73 | TRANSFER LEARNING FOR GENERALIZING A HYBRID
AUTOENCODER -ISOLATION FOREST MODEL FOR TIME SERIES
ANOMALY DETECTION IN ARRONAX
CYCLOTRON OPERATIONAL DATA*
F. Basbous†,1,2, F. Poirier1,3, F. Haddad1,2,3, D. Mateus4
IP ARRONAX, Saint -Herblain, France
2Nantes Université, Nantes, France
3Centre National de la Re... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 0 |
a4ce8174-5a7d-43be-8721-858a54750c51 | a hybrid model combining an autoencoder and IF, using
time series data from the proton beam intensity on target.
This approach significantly improved the detection of both
global and local anomalies, with no false alarms observed
during evaluation. Building on these results, the present
study investigates the use ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 0 |
dd61d9f5-7dc4-4723-b874-f92dcf05e838 | posed of complex and interconnected components and sub-systems. Therefore, anomaly detection has become an im-
portant research area, as deviations from normal opera-
tional behavior can indicate potential faults and help pre-
vent failures, irreversible damage, and costly repairs [1-4],
In our previous work [5], we... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 0 |
8efdd731-ae84-45fc-a329-27ca6eaabf41 | the model enables the detection of both global anomalies and subtle local anomalies occurring near the mean of nor-mal data. The proposed approach was shown to outperform
two alternative IF -based methods applied directly to the
raw signal and to the PCA -transformed intensity space.
Despite these promising results... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 0 |
848ac8ca-3b97-433a-9cdd-38a68a72f88f | bles the reuse of knowledge learned from a source dataset
when training models for related target datasets [6].
TL has demonstrated significant benefits in anomaly de-
tection tasks, particularly in industrial time -series applica-
tions where labelled data are limited, enhancing detection
robustness while reducing... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 0 |
39c7661b-18e9-4afc-b388-91851ef8e719 | ness [10]. However, to the best of our knowledge, the ap-
plication of TL to anomaly detection in accelerator opera-tional time -series data remains limited.
TL can be categorized according to the similarity of tasks
and domains between the source and target problems. In
general, TL is divided into inductive, trans... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 0 |
ff034991-1f24-4d26-8d56-7c1c9fda3b42 | are unlabe led.
Among these categories, transductive TL is particularly
relevant to our study since the anomaly detection task re-
mains unchanged while the data originate from different
sensors. In this work, we propose a TL framework that re-uses the latent representation learned from normal proton ____________... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 0 |
c190bbe6-9474-403b-82cc-8df543536695 | ITMO Cancer_18011 (SIRIC ILIAD).
†basbous@arronax -nantes.fr
17th International Particle Accelerator Conference, Deauville, France JACoW Publishing
ISBN: 978-3-95450-252-3 ISSN: 2673-5350 doi: 10.18429/JACoW-IPAC2026-WEP6002
2604 MC6.D13: Instrumentation: Artificial IntelligenceWEP6002 WEP: Poster session
Content ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 0 |
6f49d3d9-3f18-4c9f-9c63-621558ae6590 | beam intensity data to generalize anomaly detection across
multiple PVs. By freezing the pretrained encoder and the
latent representation and adapting only the output layer, the
proposed approach reduces training complexity while
maintaining robust detecti on of both global and local
anomalies across several PVs d... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 1 |
841c930c-ed87-4b09-ac5f-538982cd346f | tion. Only two PVs are shown for illustration.
Baseline AE -IF Architecture
The baseline model corresponds to the hybrid AE -IF ar-
chitecture introduced in our previous work [5]. The train-
ing set consists of 𝑁𝑁 vectors 𝑥𝑥𝑖𝑖∈ℝᵅ, denoted 𝑋𝑋={𝑥𝑥𝑖𝑖}𝑖𝑖=1𝑁𝑁,
where each vector represents a standardize... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 1 |
50237b1d-bbc8-44c3-938b-d8b53d778747 | 𝑥𝑥ᵢ is computed as defined in Eq. (1)
𝑀𝑀𝑀𝑀𝑀𝑀 (𝑥𝑥ᵢ,𝑥𝑥ᵢ�)=1
𝑎𝑎∑�𝑥𝑥𝑖𝑖𝑖𝑖−𝑥𝑥�𝑖𝑖𝑖𝑖�3𝑎𝑎
𝑖𝑖=1 . (1)
These reconstruction -error features are then used as input
to the IF (Fig. 1(c)), which assigns anomaly scores based
on the average path length in random tr... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 1 |
805e1a5c-261b-42b4-9527-19965e1acd36 | gle dense output layer, as illustrated in Fig. 1(b). Within the
proposed transductive TL framework, this output layer is
trained separately f or each target PV using its correspond-
ing training subset, filtered according to normal data iden-tified in the proton beam intensity on target training set.
To ensure stab... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 1 |
2ccd9420-9963-4c41-a0ab-e3e6cc3cb7a1 | 𝐾𝐾 the total number of samples in the training batch. Anom-
aly detection is then performed using IF, retrained for each
PV as in the baseline AE -IF framework.
Operational Definition of Anomalies
Despite differences in domain and scale across PVs, the
definition of normal and anomalous behavior remains un-
ch... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 1 |
839b6f32-3d6b-46ec-af2e-0ac5c4afcf19 | the defined interval. In contrast, local anomalies are more
subtle. In this case, all observations remain within the in-
terval [𝑆𝑆𝑙𝑙𝑙𝑙𝑛𝑛,𝑆𝑆ℎ𝑖𝑖𝑖𝑖ℎ], while the internal variability within the
time window is abnormally high and exceeds a predefined threshold 𝛼𝛼.
RESULTS AND DISCUSSION
Table 1: Valid... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 1 |
391316ae-c556-4032-b4a5-b01a1700311a | ISBN: 978-3-95450-252-3 ISSN: 2673-5350 doi: 10.18429/JACoW-IPAC2026-WEP6002
MC6.D13: Instrumentation: Artificial Intelligence 2605WEP: Poster session WEP6002
Content from this work may be used under the terms of the CC BY 4.0 licence (© 2026). Any distribution of this work must maintain attribution to the author(s), t... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 1 |
f862786d-36e3-4728-af4a-5b3a7f9d1ba6 | Experimental Setup
To evaluate the performance of the proposed TL frame-
work, 26 time -series datasets were used in this study, each
corresponding to a run of approximately 7 days of contin-
uous operation and including 8 representative PVs from the
C70XP cyclotron. These PVs originate from multiple sub-
systems, ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 2 |
09758465-5989-48a3-8a9f-fefbcfa1bd13 | erage of 74,000 sequences per run for each PV .
For model development, 60% of the runs (15 runs) were
allocated to training and validation, while the remaining
40% (11 unseen runs) were reserved for final testing. The anomaly rate across the PV datasets ranges from approxi-
mately 0.1% to 16%, reflecting high to mo... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 2 |
2a790fe8-f5ae-43eb-b6f7-5bd1b3e80cd2 | framework across PVs on unseen runs (mean ± std).
Validation Results
To assess whether setting the IF contamination parame-
ter equal to the anomaly rate of the training subset remains
appropriate under varying anomaly rates, a 10 -fold cross -
validation across runs [5] was performed. Two TL config-
urations wer... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 2 |
dff21e23-75e1-474e-a795-958bfa350267 | tuning configuration yields lower F1 -scores (F1 < 0.5).
Based on these results, the proposed framework was re-tained, and the IF contamination parameter set to the anom-
aly rate of the training subset remains effective across most PVs despit e variations between runs, while RF FP exhibits
the lowest F1 -score , hig... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 2 |
c7573759-ec2a-4348-9dd3-c998564aaed1 | model detects both global and local anomalies while cor-
rectly classifies normal sequences. Among the studied
PVs, RF FP achieved the lowest F1 -score, consistent with
the validation results. IV shows the second lowest perfor-
mance, which appear s to be unrelated to the decision
threshold or to the model itself ,... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 2 |
5164d1ce-a5aa-4d4a-a1ad-79cb625f6b20 | interval. As a result, stable sequences that are anomalous
with respect to run -specific thresholds may still fall within
the training range , leading to misclassification and lower
recall. This scaling limitation also explains the noticeable
variability in recall observed for IDV.
CONCLUSION
The proposed TL fra... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 2 |
3849022e-b1a2-405e-938b-5f7bfaa55167 | ISBN: 978-3-95450-252-3 ISSN: 2673-5350 doi: 10.18429/JACoW-IPAC2026-WEP6002
2606 MC6.D13: Instrumentation: Artificial IntelligenceWEP6002 WEP: Poster session
Content from this work may be used under the terms of the CC BY 4.0 licence (© 2026). Any distribution of this work must maintain attribution to the author(s), t... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 2 |
5947f77b-f9ed-4009-86d3-92fa3a07ad8b | REFERENCES
[1] F. Poirier, D. Mateus, J. Rioult, and C. Lassalle, “First
anomalies exploration from data mining and machine
learning at the ARRONAX cyclotron C70XP”, in Proc.
IPAC’23, Venice, Italy, May 2023, pp. 2273-2276.
doi:10.18429/JACoW- IPAC2023- TUPM036
[2] A. Ghribi et al., “Artificial intelligence f... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 3 |
164b6690-b028-42ac-ba62-dbca79ef1709 | doi:10.1103/PhysRevAccelBeams.27.063201
[5] F. Basbous, F. Poirier, D. Mateus, and F. Haddad, “Hybrid
autoencoder-isolation forest approach for time series
anomaly detection in C70XP cyclotron operational data at ARRONAX”, in Proc. CYC2025, Oct. 2025.
doi:10.48550/arXiv.2603.20335
[6] S. J. Pan and Q. Yang, ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 3 |
aa932c32-0d84-4493-a123-554056579bc4 | consumption in school buildings: A case of northern Swe-
den”, Energy Build. , vol. 346, p. 116129, Nov. 2025.
doi:10.1016/j.enbuild.2025.116129
[9] M. Rezakhani, T. Seyfi, and F. Afghah, “A transfer learn-
ing framework for anomaly detection in multivariate IoT
traffic data”, in Proc. IEEE ICC’25, Jun. 2025, pp. ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 3 |
b1312fd3-e9dd-44ed-b4b8-94adc868569d | optimization”, Jan. 2017, arXiv:1412.6980
doi:10.48550/arXiv.1412.6980
17th International Particle Accelerator Conference, Deauville, France JACoW Publishing
ISBN: 978-3-95450-252-3 ISSN: 2673-5350 doi: 10.18429/JACoW-IPAC2026-WEP6002
MC6.D13: Instrumentation: Artificial Intelligence 2607WEP: Poster session WEP... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6002.pdf | 3 |
b8de52cc-83cc-497f-8d3a-0a6fc0d9c3ec | GRADIENT-BASED LASER CONTROL FOR END-TO-END
PHOTOINJECTOR EMITTANCE OPTIMIZATION AT EUXFEL
D. Ilia1,2∗A. Aksoy3, Z. Amirkhanyan3, F. Brinker1, M. Cai1,2, M. E. Castro Carballo1, Y. Chen1,
J. Good1, M. Gross1, U. Große-Wortmann1, M. Guetg1, I. Hartl1, W. Hillert2, Y. Jiang1,
M. Kerstan1, A. Klemps4, M. Krasilnikov1, J. ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 0 |
06a8d3c5-ebb8-4e2b-b88f-4455f99f336a | for meeting the performance targets of the European XFEL,
particularlyforhighphotonenergiesandfuturehigh-duty-
cycle operation. Both the temporal structure of the drive-
laserpulseandtheRF-gunsettingscontributesignificantlyto
the final beam quality, yet their optimization is complicated
bystrongnonlinearitiesinthelaser... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 0 |
03d29b04-08cd-4eee-8b0b-5a1b44d10478 | ing parameters. Together, these tools provide an end-to-end
control framework for emittance reduction at EuXFEL. Ini-
tialresultsdemonstratethatthedifferentiablemodelenables
accurate temporal UV pulse shaping at EuXFEL. Work is
ongoing to integrate this approach with ML-assisted gun
optimization within the proposed end... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 0 |
330a1187-c91a-417d-9f65-2b28f708777d | bunchemittance,anditcanbemitigatedbyshapinginspace
and time the photoinjector laser pulses to control the initial
electronbunchdistribution. Inpreviouswork[ 2,3],were-
ported on NEPAL, the photoinjector laser serving EuXFEL,
FLASH,andPITZ,whichsupportsbothtemporalandspa-
tial shaping. In particular, we demonstrated UV ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 0 |
969688b5-b15c-40ca-927a-f1dda85787c1 | cultyistheopticalchainmappingSLMsettingstotheUV
profile,whichcontainsseveralstronglynonlinearstagesand
admitsnodirectanalyticalinversion. Asaresult,themanual
scheme is too slow and irreproducible to support systematic
emittancestudies,letaloneend-to-endoptimizationofthe
gunparameters. Inthispaperweaddressthisbyintroduc... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 0 |
06ea951b-478f-4a03-bd71-75b5263cab8b | optimization methods targeting the gun parameters.
PHOTOCATHODE LASER TEMPORAL
SHAPING
NEPALdeliverspicoseconddeep-UVpulsesat257nm
tothephotocathodethroughachirped-pulseamplification
(CPA)scheme followedbytwostages of frequencyconver-
sion (Fig. 1(a), top). Temporal pulse shaping is enabled
byafiber-coupledspatiallight... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 0 |
ba84af83-3e3d-428d-9686-341bc3a05841 | that the resulting intensity profile 𝐼laser(𝑡)matches a given17th International Particle Accelerator Conference, Deauville, France JACoW Publishing
ISBN: 978-3-95450-252-3 ISSN: 2673-5350 doi: 10.18429/JACoW-IPAC2026-WEP6006
2608 MC6.D13: Instrumentation: Artificial IntelligenceWEP6006 WEP: Poster session
Content fro... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 0 |
15163711-afea-4fe8-af7e-0a117d1f11f5 | Oscillator Stretcher SLMFiber
AmplifierSolid-State
AmplifierCompressor LBO BBO
Cross-
correlatordelay τLaser system
Control optimization Model fittingDifferentiable modelControl input
u(ν)
updated modelIntensity profile
I(t)(a)
−10 0 10
Time (ps)Intensity (a.u.)NRMSE = 1.70%
−10 0 10
Time (ps)NRMSE = 2.58%Target Optimi... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 1 |
12d205a5-8328-4108-842b-08dd4a9f6402 | measurementcyclestoconvergeonasingleprofile,which
is incompatible with routine photoinjector operation and
would prevent systematic emittance studies.
Weinsteadintroduceadifferentiablemodelofthelaser
system,𝑓(𝑢,𝜃) = 𝐼 model(𝑡), on which gradient descent in
𝑢efficiently locates the optimal filter; we refer to this... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 1 |
60df51f8-7c6a-41df-954f-23b809396a8e | in a small fraction of the iterations required by black-box
methods. WehaveappliedthefullschemeatEuXFEL:as
shown in Fig. 1(b), the method reproduces the target pro-
fileswithanormalizedRMSerrorbelow5%andconverges
inabout50iterations—morethananorderofmagnitude
fewerthantypicalblack-boxapproachesatcomparableac-
curacy. I... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 1 |
91fdb8a9-496e-4b44-b37f-60a4bdd48f5d | longitudinalelectrondistributionatthegun,andtherefore
over the space-charge contribution to the emittance. On the
otherhand,itenablesoperationalmodesthatrelyonatai-
lored pulse shape, including schemes recently proposed forplasma-wakefield acceleration and attosecond XFEL pulses
[5–7]. Belowwepresenttworecentapplicatio... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 1 |
c5420b6a-ad30-4722-b45d-bd6da20af9c8 | makes twin-bunch operation practical, in which both lasers
areusedtogenerateinterleavedelectronbunchesthatgen-
erate self-amplified spontaneous emission (SASE) at the
undulators (Fig. 2(b)).
(a)
(b)
Figure 2: (a) Matched temporal profiles of the two photoin-
jector lasers at EuXFEL. (b) Twin-bunch operation enabled
by ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 1 |
560f95db-4767-4fee-a538-13b3c00fd9f4 | 0.7 0.8 0.9 1.0
Laser beam diameter (mm)0.50.60.70.80.9Proj. n,xy (mm mrad)
a) Gaussian 12 ps
Gaussian 8 ps
Flattop 20 ps
Flattop 16 ps
Parabolic 20 ps
Parabolic 16 ps
25
0 25
Intensityb)
0.5
0.0 0.5
x (mm)0.5
0.00.5
y (mm)e)
25
0 25
Time (ps)c)
25
0 25
Time (ps)
Intensityd)Figure3: Transverseemittancedependenceonp... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 2 |
523a1ea6-06cc-42d8-93be-d9b08c158fbe | optimization program for EuXFEL, we carried out a ded-
icated study at the Photo Injector Test Facility at DESY
Zeuthen(PITZ)toevaluatethisdependencesystematically.
Although the full analysis is still ongoing, we report here
preliminaryresultsonthetransverseemittanceasafunction
of both the temporal profile and the lase... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 2 |
97c60589-9db1-44e8-ad0c-f42ee102a617 | than parabolic pulses of equal FWHM duration. While a
complete analysis of the data and associated uncertainties is
stillpending,re-measurementofthesameoperatingpoint
one week later returned a consistent value, hinting at a pos-
sible benefit of flat-top shaping [ 8].
CONCLUSIONS
Wehavepresentedanadaptive,gradient-base... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 2 |
c5846ece-2b17-4d67-a929-87fef43c27f2 | itytoinvestigatehowthetransverseemittancedependson
the initial laser distribution. A preliminary study at PITZ as
a function of pulse profile and spot size suggests a possible
advantageofflat-topoverparabolicshapingatequalFWHMand provides a first indication of how the temporal profile
contributes to the emittance. Ongo... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 2 |
9c1fdde1-ea58-484a-a4bb-fa578262ec3e | REFERENCES
[1]Y. Chenet al., “Characterization of low-emittance electron
beams generated by a new photocathode drive laser system
NEPAL at the European XFEL”, in Proc. IPAC’24 , Nashville,
TN,USA,pp. 388–391, Jul. 2024.
doi:1/zero.alt4.18429/JACoW-IPAC2/zero.alt424-MOPG47
[2]C.Mahnke etal.,“Novelphotocathodelasersforth... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 2 |
00eccca3-027f-4fce-945d-7cb6caa04407 | femtosecondpulseshaping”, JOSAB,vol.15,no.5,pp.1615–
1619, May 1998. doi:1/zero.alt4.1364/JOSAB.15./zero.alt4/zero.alt41615
[5]I. Kuzmin etal., “Shaping triangular picosecond laser pulses
for electron photoinjectors”, Laser Physics Letters , vol. 16,
no. 1, p. 015001, Jan. 2019.
doi:1/zero.alt4.1/zero.alt488/1612-2/zer... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 2 |
d181540d-f5bb-4ccf-b4a7-01a447f2f6c2 | ISBN: 978-3-95450-252-3 ISSN: 2673-5350 doi: 10.18429/JACoW-IPAC2026-WEP6006
2610 MC6.D13: Instrumentation: Artificial IntelligenceWEP6006 WEP: Poster session
Content from this work may be used under the terms of the CC BY 4.0 licence (© 2026). Any distribution of this work must maintain attribution to the author(s), t... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 2 |
715121c8-b980-4c80-9253-4bd17840b1db | [7]G. Ha, K. -J. Kim, J. G. Power, Y. Sun, and P. Piot, “Bunch
shaping in electron linear accelerators”, Reviews of Modern
Physics, vol. 94, no. 2, p. 025006, May 2022.
doi:1/zero.alt4.11/zero.alt43/RevModPhys.94./zero.alt425/zero.alt4/zero.alt46[8]M. Krasilnikov et al., “Experimentally minimized beam emit-
tancefroman... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6006.pdf | 3 |
716f4f7a-d549-4217-8830-0375c13edc96 | EXPLORING CONVOLUTIONAL NEURAL NETWORKS TRAINING
STRATEGIES FOR PILE UP CORRECTION IN SINGLE PARTICLE
COUNTING
M. Hamdan∗, Unaffiliated Researcher, Frankfurt, Germany
T. Habermann, Fulda University of Applied Sciences, Fulda, Germany
Abstract
Thisworkinvestigatestheeffectofdifferenttrainingstrate-
giesontheperformanceo... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 0 |
0f82c203-b4a4-403c-8649-cdb097c449e6 | toadaptcommonmetricsforprecisionandrecalltoallow
a user-defined localization tolerance. Using a dataset of
approximately 26,000 manually labeled pulses, we analyze
how these training choices impact particle counting and lo-
calization performance. The findings provide insights for
developing particlecounting systems su... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 0 |
4e04bbde-8267-4cbc-8bc3-2087b7d97b4c | shorttimeintervalsandproducesuperimposedpulseshapes
atthedetectoroutput. Typically,thenumberofpile-upsin
extracted beams from particle accelerators is much more
than one expects from a random discrete process or Poisson
process [1,2].
Previouswork[3]demonstratedhowconvolutionalneu-
ralnetworks(CNNs)canbeusedforpile-upc... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 0 |
6c0e293e-b4e7-498a-bfcb-019edba2a602 | However,whilethenetworkarchitecturewasoptimized
for deployment constraints, the influence of the training
∗Mahdie.Hamdan@outlook.comstrategy on detection behavior was not systematically inves-
tigated. In particular, because particle counting involves
sparseeventlocalizationinhighlyimbalancedtime-series
data, the loss ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 0 |
bea32080-5302-41f7-84c2-0ad91d2e4fea | particlecountingsystems andshowhowtoimprovethe ap-
proach for localization and particle counting.
BACKGROUND
TheCNNmodelpresentedin[3]isprocessingwindows
of data in sequential order. Each window of data contains
256datapoints. Figure1showsawindowofdatatogether
0 10 20 30 40 50 60 70 80
ns0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 0 |
8ad074b5-f681-47dc-a3ac-032dc09810c5 | twofeatures,whicharetheconfidencescoreandtherelative
positioninthecell. Bothfeaturesarenormalizedbetween
zero and one using a sigmoid activation function. A peak
is predicted when the confidence score is higher than 0.5.
ThisconceptissimilartoYOLO[5]boxdetectionmodels.
TheCNNmodelarchitecturefrom[3]isasimpleCNNwith
jus... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 0 |
8342636c-0929-4d86-bf1e-5b5b6888dfde | can be seen that the stride settings of the first two layers de-
fine the cell size, and the kernel size of the following layers
definehowmanyneighboringcellseachcellcantakeinto
consideration for peak prediction.
Table 1: Model Architecture for 𝑛Cells Where Each Cell
Can Detect up to𝑚Pulses
Layer Kernel Stride Filter... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 1 |
86a59deb-3de7-46b4-82c9-3683f3d84123 | established remedy, it is static — its weighting does not
adapt as training progresses — and it operates purely point-
wise, without any notion of global detection overlap. We
therefore consider theGradient Harmonizing Mechanism
(GHM) [6], which remedies the static nature of focal loss
by adaptively reweighting example... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 1 |
5b4260a3-3e2f-49a5-9cfd-8b5cbe5e494b | 𝑔over a training batch is highly non-uniform. Easy nega-
tives cluster densely near 𝑔≈0and, despite contributing
littleindividually,overwhelmthegradientinaggregate. A
smaller cluster of hard outliers near 𝑔≈1can further desta-
bilizetraining. Thefocallosssuppresses theeasymajority,
but its weighting is static and ig... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 1 |
b459deb7-4820-4970-a176-7ee55af75a44 | 𝐿GHM-C =1
𝑁𝑁∑︁
𝑖=1𝑁
𝑅𝑗(𝑖)𝐿𝐶𝐸(𝑝𝑖,𝑝∗
𝑖),
where𝑗(𝑖)denotes the bin of example 𝑖. The weights are
normalized per batch for stability. Crucially, unlike focal
loss, GHM isadaptive: the bin counts 𝑅𝑗are recomputed
eachmini-batch,sotheweightingautomaticallytracksthe
evolving training distribution without a... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 1 |
848fd6cc-d973-4797-aeb2-1bc76295e917 | where𝑝𝑡=𝑝if𝑝∗=1and𝑝𝑡=1−𝑝otherwise, the
region-based counterpart is derived by considering the Dice
loss,
L𝐷𝑆𝐶=1−2𝑇𝑃
2𝑇𝑃+𝐹𝑃+𝐹𝑁,
where𝑇𝑃,𝐹𝑃,𝐹𝑁are true positives, false positives, and
false negatives. Since Dice treats FP and FN symmetrically,
theTversky indexgeneralizes it with separate weights,
�... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 1 |
79103d63-50d8-4e29-a54d-a080c58c006c | focal exponents of the two terms to a shared 𝛾(as(1−𝛾)
and𝛾respectively)andreusing 𝛿fromEq. (2)astheclass-
weighting parameter in Eq. (1):
L𝑈𝐹=−𝜆𝛿(1−𝑝 𝑡)1−𝛾log(𝑝𝑡)
+(1−𝜆)
1−𝑇𝑃
𝑇𝑃+𝛿𝐹𝑃+(1−𝛿)𝐹𝑁𝛾
.
Here𝛿controlstheFP/FNtradeoff, 𝛾governsfocalmodu-
lationstrengthforbothcomponentssimultaneously,a... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 1 |
1ed8f2f3-58e2-47b1-a0b5-132269fae799 | METRICS
In [3], the absolute difference between the number of
predictedpeaksandthenumberoftruepeakswastakento
measure the quality of the approach.
However,sincethemodelpredictspeaksonaper-window
basis, this metric fails to penalize under- or overcounting
within a given window. To address this limitation, we intro-
duce... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 2 |
bce74bb4-c4b5-438f-882d-81441d3c0c0b | tomappredictedpeaklocationsandactualpeaklocations
ontoa1Dimageofacertainresolution. Forexample,witha
windowsizeof8,whenthemodelpredictsthreepeaksatpo-
sition 0,4,4andtheactualpeaksareat 4,5,thepredictedand
actualpeakscanbemappedontoa1Dpixelimage 𝑝map=
(1,0,0,0,2,0,0,0) and𝑦map=(0,0,0,0,1,1,0,0) . This
allows counting... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 2 |
ad18c3cb-ac92-4f03-90ea-dee69d281e73 | guitywithinagivenerrorrange. Also,asthismetricrelies
on the confidence score and relative position predictions
of the model, the metric provides a more comprehensive
performance overview of the model. We then take the com-
monly used F1 score [8], which combines precision and
recall, to measure localization performance... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 2 |
485c2b97-61b4-4ccc-924e-708ba66af4be | functionareusedasareference,asitwasthelossfunction
used in [3].Particle Counting
For particle counting, we searched for the best 𝑆𝐶score.
Table 2 shows the best results for particle counting for each
loss function. The best hyperparameters for focal loss were
Table 2: The Best Results for Particle Counting
LossF 1(0)... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 2 |
becf8bee-07c1-4971-8a18-8117a878eab6 | function. The best hyperparameters for focal loss were
Table 3: The Best Results for Particle Localization
LossF 1(0)F 1(1)SC Diff.
Focal Loss [3]0.9401 0.990497.37% 301
GHM Loss 0.9357 0.987 97.55% -445
Uni. F. Loss 0.9388 0.9887 97.11% 573
𝑐=0.5,𝛼=0.5,𝛾=3.0 , for GHM loss 𝑐=0.03125 , and
forunifiedfocalloss 𝑐=0.... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 2 |
e65fb02f-9f86-4547-9082-2e4c34020ba4 | CNNsusedforpile-upcorrectioninsingleparticlecounting.
We determined the most fitting hyperparameter settings for
eachloss functionto tunetheCNN forprioritizing particle
counting or particle localization. We also proposed new
metrics to measure particle counting and localization per-
formance. It was shown that the unif... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 2 |
17f96b59-e8bb-4c01-8fb1-590d9f387881 | ISBN: 978-3-95450-252-3 ISSN: 2673-5350 doi: 10.18429/JACoW-IPAC2026-WEP6009
2614 MC6.D13: Instrumentation: Artificial IntelligenceWEP6009 WEP: Poster session
Content from this work may be used under the terms of the CC BY 4.0 licence (© 2026). Any distribution of this work must maintain attribution to the author(s), t... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 2 |
f511a1a4-7e6f-4b17-b94c-cdf5a22ce80e | REFERENCES
[1]R. Singh, P. Forck, and S. Sorge, “Reducing fluctuations in
slow-extractionbeamspillusingtransit-time-dependenttune
modulation”,Phys.Rev.Appl.,vol.13,no.4,p.044076,Apr.
2020.
doi:\\10.1103/PhysRevApplied.13.044076
[2]P. Niedermayer and R. Singh, “Excitation signal optimization
for minimizing fluctuations ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 3 |
64d8dda6-9ac9-443a-b031-b2638b797a2a | ConferenceonComputerVision(ICCV),pp.2999–3007,2017.
doi:\\10.1109/ICCV.2017.324[5]J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You
onlylookonce:unified,real-timeobjectdetection”,in2016
IEEEConferenceonComputerVisionandPatternRecognition
(CVPR), pp. 779–788, 2016.
doi:\\10.1109/CVPR.2016.91
[6]B.Li,Y.Liu,andX.Wa... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 3 |
3f09f52e-3bbb-48d4-985f-072158db2b79 | [8]M.SokolovaandG.Lapalme,“Asystematicanalysisofper-
formance measures for classification tasks”,Information Pro-
cessing and Management, vol. 45, no. 4, pp. 427–437, 2009.
doi:\\10.1016/j.ipm.2009.03.00217th International Particle Accelerator Conference, Deauville, France JACoW Publishing
ISBN: 978-3-95450-252-3 ISSN:... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6009.pdf | 3 |
820e3506-7184-45ff-ae88-2765858d856b | FEL POWER PROFILE PREDICTIONS WITH IMAGE-BASED MACHINE
LEARNING AT FLASH
J. Lundquist∗, S. Werin and F. Curbis, Dept. of Physics, Lund University, Lund, Sweden
E. Mansten, MAX IV Laboratory, Lund, Sweden
G. Goetzke, Deutsches Elektronen-Synchrotron DESY, Hamburg, Germany
Abstract
Accelerator operators and beamline user... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 0 |
1ebb3169-504a-4b04-b8ae-38d8868643f6 | tudinal phase-space (LPS) of electron beams unaffected by
the FEL process, originally measured using a transverse de-
flectingstructure. Thepredictionswereusedincombination
withlongitudinalmeasurementsoftheLPSoftheelectron
beam after lasing, which does not interfere with delivery
to users, to reconstruct the FEL pulse.... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 0 |
9b41c042-9217-477f-ac60-9fe0a58f282f | time. Applications more suited to routine operations and
online monitoring are in high demand [2].
When using the electron beam for FEL power profile
measurements, one requires full longitudinal phase space
(LPS)measurements,whichisdoneatFLASHusingatrans-
verse deflecting structure (TDS) [3,4]. TDS measurements
can be ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 0 |
912b6cde-b562-43ab-9ba4-837346cef751 | livernon-destructive,single-shotprofiles,suchasTHz[5]or
angular streaking [6], can be considered experiments them-
selves and often require extensive use of dedicated machine
time.
∗johan.lundquist@maxiv.lu.seThere is a high demand for applications more suited to
routine operations and online monitoring. Advancements i... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 0 |
915e1915-a7c0-4101-b63e-02a03fcbc5f3 | profile, we implement an image-based virtual diagnostic
topredictthebeam’sfullLPSdistribution,beforecompar-
ing with measurements during lasing for FEL power profile
reconstruction. This is also compared directly with a non-
machine learning reference method for reconstructing the
power profiles.
EXPERIMENTAL SETUP
FLA... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 0 |
f1027fae-6c8e-4907-a775-fc78fdd29d24 | ing structure for beam streaking [8,9]. The 0.8m PolariX
structure is set up 2m after the last FLASH2 undulator. Af-
ter the RF streaking, the electron beam is diverted from the
photonbeampathbya3.5◦-bendmagnetandduringnormal
delivery, one bunch in the bunch train is deflected to a scin-
tillating screen for imaging. F... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 0 |
8e34662c-c66e-444b-b08e-215fc90ae736 | on the screen respectively.
The data used in this study consisted of 5919 images
collectedwithoutthebeamlasingand3146imagescollected
during lasing, all collected at a nominal setpoint for the
FLASHaccelerator. Lasingwascontrolledwithahorizontal17th International Particle Accelerator Conference, Deauville, France JACoW ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 0 |
8444b838-5249-4481-a55c-4f335db84485 | 0.3 0.4 0.5 0.6 0.7 0.8 0.9
R2 Score0.02.55.07.510.012.5CountsR2 Score Distribution
N = 91
Mean = 0.792
Median = 0.844Figure 1: Coefficient of determination, 𝑅2-score, of the VD
predictions across the test dataset with 7 cases of 𝑅2<0
removed.
kick to the beam trajectory, applied using steering magnets
before the und... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 1 |
55a19160-a1de-4097-bf1d-b717d657e27a | of1.3milliontrainableparameters. TheANNwastrained
using the ADAM optimizer and with mean absolute error
(MAE)asthelossfunction. ThetrainedVDachievedafinal
mediancoefficientofdeterminationof84.4 ±13.8%onatest
dataset of 91 non-lasing images, randomly selected from
the full non-lasing dataset. This smaller test dataset h... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 1 |
33739dff-4f37-445c-b7d1-9cf213962f9d | setups.
FEL Power Profile Reconstruction
ThemethodologyforreconstructingFELpowerprofiles
followed similar steps as outlined in [2]. The central energy
wascalculatedforeachslicealongthelongitudinalcoordi-
nate for the lasing and non-lasing LPS distributions. The
lasing energy profile is then subtracted from the non-lasi... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 1 |
3ecc2a48-b512-4aa4-b79c-c06779283a37 | tions were collected, clustered into groups, and the average
properties of these non-lasing data are compared to those
lasing. The group with the highest similarity for each lasing
profile was chosen for the FEL power profile reconstruction.
In order to demonstrate the accuracy of the predicted pro-
files, we compare t... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 1 |
3924c479-e1ef-4bcf-82bd-a897c20a51c9 | reconstruction can be seen in Fig. 2. In the last plot, the
two FEL power profile reconstructions are shown, using
eithertheVD-orcorrelation-basednon-lasingLPSforthe
comparative reconstruction. The specific case shown in this
figure represents the median case in terms of correlation
between the two reconstructed FEL po... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 1 |
52e4e2c1-394b-40e6-94b1-d3f95f229e2e | pre-processing.
Figure2showsonlyasinglecaseofthedifferentrecon-
structions,andforacleareroverviewofalltheFELprofiles,
onecanlookatthemeanprofiles,calculatedbasedoneachof
the FEL profiles reconstructed using both techniques. This
isshowninFig.3,alongwithshadedregionsforonestan-
dard deviation of the distribution of powe... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 1 |
c25268d9-c739-4735-9aa5-9e0953efe04e | CONCLUSION AND OUTLOOK
Thefinalresultsofthisprojectdemonstratethefeasibility
of leveraging ML-based virtual diagnostics trained on non-
destructive beam data for FEL power profile reconstruction.
While such approaches have been tested in the past, here we
show the use of full non-lasing LPS predictions to enable
non-in... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 1 |
e13bd251-8269-4a7c-bdc1-63997503f4d8 | -1.4
0.0
1.3dE/E [%]Lasing Shot
-1.4
0.0
1.3dE/E [%]Non-lasing - VD Prediction
-245.0 -122.0 0.0 122.0 245.0
t [fs]024Power Normalized
to GMD [GW]
FEL Power Profile Reconstructions
VD-based
Correlation-basedFigure 2: The LPS distribution of a lasing shot and its corresponding non-lasing shot based on a VD prediction. ... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 2 |
2d11b004-4bde-4db9-967e-4fb104ff238c | based on both virtual diagnostic predictions and correlation
selectionofnon-lasingshots. Plottedwithshadedregions
representing±1𝜎.
with a correlation-based reference method the VD-based
approachshowsagoodoverallagreement,suggestingthat
theVDiscapableofcapturingthekeyfeaturesrelevantto
FEL power.
Therearesomedrawbackst... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 2 |
780faa02-715b-4948-8ca4-19b314ddfb2a | alternativeto traditionalmethodswhich couldbemodified
andimprovedforonlinemonitoringoftheFELpowerprofile.
The few discrepancies present in the results could likely be
alleviated with fine tuning of the models.
Future work could focus on improved robustness of the
VD model, for example with improved network architecture... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 2 |
25fea3df-b635-4599-80e8-c5f0269df80a | implementation in control rooms.
ACKNOWLEDGEMENTS
The network training was enabled by resources provided
by LUNARC, The Centre for Scientific and Technical Com-
putingatLundUniversity. Thisresearchwasalsosupported
inpartthroughtheMaxwell computationalresourcesoper-
ated at DESY, Hamburg, Germany.
REFERENCES
[1]Z.Huanga... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 2 |
0c478f88-c448-4757-9dad-cddcf7b5e36a | [2]T. Korten, V. Rybnikov, P. Steinbach, and N. Mirian, “Vir-
tual pulse reconstruction diagnostic for single-shot measure-
mentoffreeelectronlaserradiationpower”,Phys.Rev.Accel.
Beams, vol. 28, no. 3, p. 030703, Mar. 2025.
doi:10.1103/PhysRevAccelBeams.28.030703
[3]W. K. H. Panofsky and W. A. Wenzel, “Some considerati... | D:\Daily_ihep\Conferences\IPAC2026\WEP\WEP6014.pdf | 3 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.