Keyword: network
Paper Title Other Keywords Page
MOPA41 Diagnostics for LINAC Optimization with Machine Learning linac, DTL, controls, diagnostics 139
 
  • R.V. Sharankova, M.W. Mwaniki, K. Seiya, M.E. Wesley
    Fermilab, Batavia, Illinois, USA
 
  The Fer­mi­lab Linac de­liv­ers 400 MeV H beam to the rest of the ac­cel­er­a­tor chain. Pro­vid­ing sta­ble in­ten­sity, en­ergy, and emit­tance is key since it di­rectly af­fects down­stream ma­chines. To op­er­ate high cur­rent beam, ac­cel­er­a­tors must min­i­mize un­con­trolled par­ti­cle loss; this is gen­er­ally ac­com­plished by min­i­miz­ing beam emit­tance. Am­bi­ent tem­per­a­ture and hu­mid­ity vari­a­tions are known to af­fect res­o­nance fre­quency of the ac­cel­er­at­ing cav­i­ties which in­duces emit­tance growth. In ad­di­tion, the en­ergy and phase space dis­tri­b­u­tion of par­ti­cles emerg­ing from the ion source are sub­ject to fluc­tu­a­tions. To counter these ef­fects we are work­ing on im­ple­ment­ing dy­namic lon­gi­tu­di­nal pa­ra­me­ter op­ti­miza­tion based on Ma­chine Learn­ing (ML). As an input for the ML model, sig­nals from beam di­ag­nos­tic have to be well un­der­stand and re­li­able. We have been re­vis­it­ing di­ag­nos­tics in the linac. In this pre­sen­ta­tion we dis­cuss the sta­tus of the di­ag­nos­tics and beam stud­ies as well as the sta­tus and plans for ML-based op­ti­miza­tion.  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-MOPA41  
About • Received ※ 05 August 2022 — Accepted ※ 06 August 2022 — Issue date ※ 07 September 2022  
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
MOPA89 RHIC Electron Beam Cooling Analysis Using Principle Component and Autoencoder Analysis luminosity, electron, ECR, beam-cooling 260
 
  • A.D. Tran, Y. Hao
    FRIB, East Lansing, Michigan, USA
  • X. Gu
    BNL, Upton, New York, USA
 
  Funding: Work supported by the US Department of Energy under contract No. DE-AC02-98CH10886.
Prin­ci­pal com­po­nent analy­sis and au­toen­coder analy­sis were used to an­a­lyze the ex­per­i­men­tal data of RHIC op­er­a­tion with low en­ergy RHIC elec­tron cool­ing (LEReC). This is un­su­per­vised learn­ing which in­cludes elec­tron beam set­tings and ob­serv­able dur­ing op­er­a­tion. Both analy­ses were used to gauge the di­men­sional re­ducibil­ity of the data and to un­der­stand which fea­tures are im­por­tant to beam cool­ing.
 
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-MOPA89  
About • Received ※ 02 August 2022 — Revised ※ 05 August 2022 — Accepted ※ 06 August 2022 — Issue date ※ 12 August 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
MOPA90 Relating Initial Distribution to Beam Loss on the Front End of a Heavy-Ion Linac Using Machine Learning simulation, LEBT, emittance, controls 263
 
  • A.D. Tran, Y. Hao
    FRIB, East Lansing, Michigan, USA
  • J.L. Martinez Marin, B. Mustapha
    ANL, Lemont, Illinois, USA
 
  Funding: This work was supported by a sub-reward from Argonne National Laboratory and supported by the U.S. Department of Energy, under Contract No. DE-AC02-06CH11357.
This work demon­strates using a Neural Net­work and a Gauss­ian Process to model the ATLAS front-end. Var­i­ous neural net­work ar­chi­tec­tures were cre­ated and trained on the ma­chine set­tings and out­puts to model the phase space pro­jec­tions. The model was then trained on a dataset, with non-lin­ear dis­tor­tion, to gauge the trans­fer­abil­ity of the model from sim­u­la­tion to ma­chine.
 
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-MOPA90  
About • Received ※ 02 August 2022 — Revised ※ 05 August 2022 — Accepted ※ 06 August 2022 — Issue date ※ 11 September 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUYE3 An Open-Source Based Data Management and Processing Framework on a Central Server for Scientific Experimental Data framework, experiment, software, data-management 307
 
  • A. Liu, J.R. Callahan, S. Poddar, W. Si
    Euclid TechLabs, Solon, Ohio, USA
  • J. Gao
    AJS Smartech LLC, Naperville, TX, USA
 
  Funding: This work is supported by the US DOE SBIR program under contract number DE-SC0021512.
The ever-ex­pand­ing size of ac­cel­er­a­tor op­er­a­tion and ex­per­i­men­tal data in­clud­ing those gen­er­ated by elec­tron mi­cro­scopes and beam­line fa­cil­i­ties ren­ders most pro­pri­etary soft­ware in­ef­fi­cient at man­ag­ing data. The Find­abil­ity, Ac­ces­si­bil­ity, In­ter­op­er­abil­ity, and Reuse (FAIR) prin­ci­ples of dig­i­tal as­sets re­quire a con­ve­nient plat­form for users to share and man­age data on. An open-source data frame­work for stor­ing raw data and meta­data, host­ing data­bases, and pro­vid­ing a plat­form for data pro­cess­ing and vi­su­al­iza­tion is highly de­sir­able. In this paper, we pre­sent an open-source, in­fra­struc­ture-in­de­pen­dent data man­age­ment soft­ware frame­work, named by Eu­clid-Nexus­LIMS, to archive, reg­is­ter, record, vi­su­al­ize and process ex­per­i­men­tal data. The soft­ware was tar­geted ini­tially for elec­tron mi­cro­scopes, but can be widely ap­plied to all sci­en­tific ex­per­i­men­tal data.
 
slides icon Slides TUYE3 [5.891 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-TUYE3  
About • Received ※ 04 August 2022 — Revised ※ 07 August 2022 — Accepted ※ 08 August 2022 — Issue date ※ 24 August 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUYE4 Machine Learning for Anomaly Detection and Classification in Particle Accelerators injection, linac, operation, controls 311
 
  • I. Lobach, M. Borland, K.C. Harkay, N. Kuklev, A. Sannibale, Y. Sun
    ANL, Lemont, Illinois, USA
 
  Funding: The work is supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, under Contract No. DE-AC02-06CH11357.
We ex­plore the pos­si­bil­ity of using a Ma­chine Learn­ing (ML) al­go­rithm to iden­tify the source of oc­ca­sional poor per­for­mance of the Par­ti­cle Ac­cu­mu­la­tor Ring (PAR) and the Linac-To-PAR (LTP) trans­port line, which are parts of the in­jec­tor com­plex of the Ad­vanced Pho­ton Source (APS) at Ar­gonne Na­tional Lab. The cause of re­duced in­jec­tion or ex­trac­tion ef­fi­cien­cies may be as sim­ple as one pa­ra­me­ter being out of range. Still, it may take an ex­pert con­sid­er­able time to no­tice it, whereas a well-trained ML model can point at it in­stantly. In ad­di­tion, a ma­chine ex­pert might not be im­me­di­ately avail­able when a prob­lem oc­curs. There­fore, we began by fo­cus­ing on such sin­gle-pa­ra­me­ter anom­alies. The train­ing data were gen­er­ated by cre­at­ing con­trolled per­tur­ba­tions of sev­eral pa­ra­me­ters of PAR and LTP one-by-one, while con­tin­u­ously log­ging all avail­able process vari­ables. Then, sev­eral ML clas­si­fiers were trained to rec­og­nize cer­tain sig­na­tures in the logged data and link them to the sources of poor ma­chine per­for­mance. Pos­si­ble ap­pli­ca­tions of au­toen­coders and vari­a­tional au­toen­coders for un­su­per­vised anom­aly de­tec­tion and for anom­aly clus­ter­ing were con­sid­ered as well.
 
slides icon Slides TUYE4 [9.534 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-TUYE4  
About • Received ※ 03 August 2022 — Revised ※ 07 August 2022 — Accepted ※ 08 August 2022 — Issue date ※ 28 August 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPA29 Machine Learning for Predicting Power Supply Trips in Storage Rings storage-ring, power-supply, quadrupole, sextupole 413
 
  • I. Lobach, M. Borland, G.I. Fystro, A. Sannibale, Y. Sun
    ANL, Lemont, Illinois, USA
  • A. Diaw, J.P. Edelen
    RadiaSoft LLC, Boulder, Colorado, USA
 
  Funding: The work is supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, under Contract No. DE-AC02-06CH11357.
In the Ad­vanced Pho­ton Source (APS) stor­age ring at Ar­gonne Na­tional Lab, trips in the mag­net power sup­plies (PSs) lead to a com­plete elec­tron beam loss a few times a year. This re­sults in un­ex­pected in­ter­rup­tions of the users’ ex­per­i­ments. In this con­tri­bu­tion, we in­ves­ti­gate the his­tor­i­cal data for the last two decades to find pre­cur­sors for the PS trips that could pro­vide an ad­vance no­tice for fu­ture trips and allow some pre­ven­tive ac­tion by the ring op­er­a­tor or by the PS main­te­nance team. Var­i­ous un­su­per­vised anom­aly de­tec­tion mod­els can be trained on the vast amounts of avail­able ref­er­ence data from the beam­time pe­ri­ods that ended with an in­ten­tional beam dump. We find that such mod­els can some­times de­tect trip pre­cur­sors in PS cur­rents, volt­ages, and in the tem­per­a­tures of mag­nets, ca­pac­i­tors and tran­sis­tors (com­po­nents of PSs).
 
poster icon Poster TUPA29 [2.116 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-TUPA29  
About • Received ※ 03 August 2022 — Revised ※ 07 August 2022 — Accepted ※ 10 August 2022 — Issue date ※ 18 August 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPA34 Model-Based Calibration of Control Parameters at the Argonne Wakefield Accelerator controls, gun, simulation, wakefield 427
 
  • I.P. Sugrue, B. Mustapha, P. Piot, J.G. Power
    ANL, Lemont, Illinois, USA
  • N. Krislock
    Northern Illinois University, DeKalb, Illinois, USA
 
  Par­ti­cle ac­cel­er­a­tors uti­lize a large num­ber of con­trol pa­ra­me­ters to gen­er­ate and ma­nip­u­late beams. Dig­i­tal mod­els and sim­u­la­tions are often used to find the best op­er­at­ing pa­ra­me­ters to achieve a set of given beam pa­ra­me­ters. Un­for­tu­nately, the op­ti­mized physics pa­ra­me­ters can­not pre­cisely be set in the con­trol sys­tem due to, e.g., cal­i­bra­tion un­cer­tain­ties. We de­vel­oped a data-dri­ven physics-in­formed sur­ro­gate model using neural net­works to re­place dig­i­tal mod­els re­ly­ing on beam-dy­nam­ics sim­u­la­tions. This sur­ro­gate model can then be used to per­form quick di­ag­nos­tics of the Ar­gonne Wake­field ac­cel­er­a­tor in real time using non­lin­ear least-squares meth­ods to find the most likely op­er­at­ing pa­ra­me­ters given a mea­sured beam dis­tri­b­u­tion.  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-TUPA34  
About • Received ※ 05 August 2022 — Revised ※ 09 August 2022 — Accepted ※ 10 August 2022 — Issue date ※ 24 September 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPA41 Applications of Machine Learning in Photo-Cathode Injectors laser, electron, controls, cathode 441
 
  • A. Aslam
    UNM-ECE, Albuquerque, USA
  • M. Babzien
    BNL, Upton, New York, USA
  • S. Biedron
    Element Aero, Chicago, USA
 
  To con­fig­ure a pho­toin­jec­tor to re­pro­duce a given elec­tron bunch with the de­sired char­ac­ter­is­tics, it is nec­es­sary to ad­just the op­er­at­ing pa­ra­me­ters with high pre­ci­sion. More or less, the fine tun­abil­ity of the laser pa­ra­me­ters are of ex­treme im­por­tance as we try to model fur­ther ap­pli­ca­tions of the pho­toin­jec­tor. The laser pulse in­ci­dent on the pho­to­cath­ode crit­i­cally af­fects the elec­tron bunch 3D phase space. Pa­ra­me­ters such as the laser pulse trans­verse shape, total en­ergy, and tem­po­ral pro­file must be con­trolled in­de­pen­dently, any laser pulse vari­a­tion over both short and long-time scales also re­quires cor­rec­tion. The abil­ity to pro­duce ar­bi­trary laser in­ten­sity dis­tri­b­u­tions en­ables bet­ter con­trol of elec­tron bunch trans­verse and lon­gi­tu­di­nal emit­tance by af­fect­ing the space-charge forces through­out the bunch. In an ac­cel­er­a­tor em­ploy­ing a pho­toin­jec­tor, elec­tron op­tics in the beam­line down­stream are used to trans­port, ma­nip­u­late, and char­ac­ter­ize the elec­tron bunch. The ad­just­ment of the elec­tron op­tics to achieve a de­sired elec­tron bunch at the in­ter­ac­tion point is a much bet­ter un­der­stood prob­lem than laser ad­just­ment, so this re­search em­pha­sizes laser shap­ing.  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-TUPA41  
About • Received ※ 30 July 2022 — Revised ※ 12 August 2022 — Accepted ※ 13 August 2022 — Issue date ※ 07 September 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPA53 Modeling of Nonlinear Beam Dynamics via a Novel Particle-Mesh Method and Surrogate Models with Symplectic Neural Networks simulation, electron, radiation, synchrotron-radiation 462
 
  • C.-K. Huang, O. Beznosov, J.W. Burby, B.E. Carlsten, G.A. Dilts, J. Domine, R. Garimella, A. Kim, T.J. Kwan, H.N. Rakotoarivelo, R.W. Robey, B. Shen, Q. Tang
    LANL, Los Alamos, New Mexico, USA
  • F.Y. Li
    New Mexico Consortium, Los Alamos, USA
 
  Funding: Work supported by the LDRD program at Los Alamos National Laboratory and the ASCR SciML program of DOE.
The self-con­sis­tent non­lin­ear dy­nam­ics of a rel­a­tivis­tic charged par­ti­cle beam, par­tic­u­larly through the in­ter­ac­tion with its com­plete self-fields, is a fun­da­men­tal prob­lem un­der­pin­ning many ac­cel­er­a­tor de­sign is­sues in high bright­ness beam ap­pli­ca­tions, as well as the de­vel­op­ment of ad­vanced ac­cel­er­a­tors. A novel self-con­sis­tent par­ti­cle-mesh code, CoSyR [1], is de­vel­oped based on a La­grangian method for the cal­cu­la­tion of the beam par­ti­cles’ ra­di­a­tion near-fields and as­so­ci­ated beam dy­nam­ics. Our re­cent sim­u­la­tions re­veal the slice emit­tance growth in a bend and com­plex in­ter­play be­tween the lon­gi­tu­di­nal and trans­verse dy­nam­ics that are not cap­tured in the 1D lon­gi­tu­di­nal sta­tic-state Co­her­ent Syn­chro­tron Ra­di­a­tion (CSR) model. We fur­ther show that sur­ro­gate mod­els with sym­plec­tic neural net­works can be trained from sim­u­la­tion data with sig­nif­i­cant time-sav­ings for the mod­el­ing of non­lin­ear beam dy­nam­ics ef­fects. Pos­si­bil­ity to ex­tend such sur­ro­gate mod­els for the study of spin-or­bital cou­pling is also briefly dis­cussed.
[1] C.-K. Huang et al., Nucl. Instruments Methods Phys. Res. Sect. A, vol. 1034, p. 166808, 2022.
 
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-TUPA53  
About • Received ※ 25 July 2022 — Revised ※ 03 August 2022 — Accepted ※ 09 August 2022 — Issue date ※ 11 August 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPA62 LANSCE Control System’s 50th Anniversary controls, EPICS, timing, data-acquisition 482
 
  • M. Pieck, C.D. Hatch, J.O. Hill, H.A. Watkins, E.E. Westbrook
    LANL, Los Alamos, New Mexico, USA
 
  After al­most ex­actly 50 years in ser­vice, the LAN­SCE (Los Alamos Neu­tron Sci­ence Cen­ter) con­trol sys­tem has achieved a major mile­stone, re­plac­ing its orig­i­nal and re­li­able RICE (Re­mote In­stru­men­ta­tion and Con­trol Equip­ment) with a mod­ern cus­tomized con­trol sys­tem. The task of re­plac­ing RICE was chal­leng­ing be­cause of its tech­nol­ogy (late 1960’s), num­ber of chan­nels (>10,000), unique char­ac­ter­is­tics (all-mod­ules data takes, timed/fla­vored data takes) and that it was de­signed as an in­te­gral part of the whole ac­cel­er­a­tor. We dis­cuss the his­tory, RICE in­te­gral ar­chi­tec­ture, up­grade ef­forts, and the new sys­tem pro­vid­ing cut­ting-edge ca­pa­bil­i­ties. The bound­ary con­di­tion was that up­grades only could be im­ple­mented dur­ing the an­nual four-month ac­cel­er­a­tor main­te­nance out­age. This led to a multi-phased pro­ject which turned out to be about an 11-year ef­fort.  
poster icon Poster TUPA62 [1.985 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-TUPA62  
About • Received ※ 02 August 2022 — Revised ※ 09 August 2022 — Accepted ※ 10 August 2022 — Issue date ※ 25 September 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
WEPA29 Real-Time Cavity Fault Prediction in CEBAF Using Deep Learning cavity, cryomodule, SRF, experiment 687
 
  • M. Rahman, K.M. Iftekharuddin
    ODU, Norfolk, Virginia, USA
  • A. Carpenter, T.S. McGuckin, C. Tennant, L.S. Vidyaratne
    JLab, Newport News, Virginia, USA
 
  Funding: Authored by Jefferson Science Associates, LLC under U.S. DOE Contract No. DE-AC05-06OR23177.
Data-dri­ven pre­dic­tion of fu­ture faults is a major re­search area for many in­dus­trial ap­pli­ca­tions. In this work, we pre­sent a new pro­ce­dure of real-time fault pre­dic­tion for su­per­con­duct­ing ra­dio-fre­quency (SRF) cav­i­ties at the Con­tin­u­ous Elec­tron Beam Ac­cel­er­a­tor Fa­cil­ity (CEBAF) using deep learn­ing. CEBAF has been af­flicted by fre­quent down­time caused by SRF cav­ity faults. We per­form fault pre­dic­tion using pre-fault RF sig­nals from C100-type cry­omod­ules. Using the pre-fault sig­nal in­for­ma­tion, the new al­go­rithm pre­dicts the type of cav­ity fault be­fore the ac­tual onset. The early pre­dic­tion may en­able po­ten­tial mit­i­ga­tion strate­gies to pre­vent the fault. In our work, we apply a two-stage fault pre­dic­tion pipeline. In the first stage, a model dis­tin­guishes be­tween faulty and nor­mal sig­nals using a U-Net deep learn­ing ar­chi­tec­ture. In the sec­ond stage of the net­work, sig­nals flagged as faulty by the first model are clas­si­fied into one of seven fault types based on learned sig­na­tures in the data. Ini­tial re­sults show that our model can suc­cess­fully pre­dict most fault types 200 ms be­fore onset. We will dis­cuss rea­sons for poor model per­for­mance on spe­cific fault types.
 
poster icon Poster WEPA29 [1.339 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-WEPA29  
About • Received ※ 02 August 2022 — Revised ※ 07 August 2022 — Accepted ※ 11 August 2022 — Issue date ※ 17 August 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
WEPA38 Progress on Machine Learning for the SNS High Voltage Converter Modulators klystron, linac, electron, simulation 715
 
  • M.I. Radaideh, S.M. Cousineau, D. Lu
    ORNL, Oak Ridge, Tennessee, USA
  • T.J. Britton, K. Rajput, M. Schram, L.S. Vidyaratne
    JLab, Newport News, Virginia, USA
  • G.C. Pappas, J.D. Walden
    ORNL RAD, Oak Ridge, Tennessee, USA
 
  The High-Volt­age Con­verter Mod­u­la­tors (HVCM) used to power the kly­strons in the Spal­la­tion Neu­tron Source (SNS) linac were se­lected as one area to ex­plore ma­chine learn­ing due to re­li­a­bil­ity is­sues in the past and the avail­abil­ity of large sets of archived wave­forms. Progress in the past two years has re­sulted in gen­er­at­ing a sig­nif­i­cant amount of sim­u­lated and mea­sured data for train­ing neural net­work mod­els such as re­cur­rent neural net­works, con­vo­lu­tional neural net­works, and vari­a­tional au­toen­coders. Ap­pli­ca­tions in anom­aly de­tec­tion, fault clas­si­fi­ca­tion, and prog­nos­tics of ca­pac­i­tor degra­da­tion were pur­sued in col­lab­o­ra­tion with the Jef­fer­son Lab­o­ra­tory, and early promis­ing re­sults were achieved. This paper will dis­cuss the progress to date and pre­sent re­sults from these ef­forts.  
poster icon Poster WEPA38 [1.320 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-WEPA38  
About • Received ※ 25 July 2022 — Revised ※ 08 August 2022 — Accepted ※ 11 August 2022 — Issue date ※ 03 October 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
WEPA40 The L-CAPE Project at FNAL controls, operation, linac, alignment 719
 
  • M. Jain, V.C. Amatya, G.U. Panapitiya, J.F. Strube
    PNNL, Richland, Washington, USA
  • B.F. Harrison, K.J. Hazelwood, W. Pellico, B.A. Schupbach, K. Seiya, J.M. St. John
    Fermilab, Batavia, Illinois, USA
 
  The con­trols sys­tem at FNAL records data asyn­chro­nously from sev­eral thou­sand Linac de­vices at their re­spec­tive ca­dences, rang­ing from 15Hz down to once per minute. In case of down­times, cur­rent op­er­a­tions are mostly re­ac­tive, in­ves­ti­gat­ing the cause of an out­age and la­bel­ing it after the fact. How­ever, as one of the most up­stream sys­tems at the FNAL ac­cel­er­a­tor com­plex, the Linac’s fore­knowl­edge of an im­pend­ing down­time as well as its du­ra­tion could prompt down­stream sys­tems to go into standby, po­ten­tially lead­ing to en­ergy sav­ings. The goals of the Linac Con­di­tion Anom­aly Pre­dic­tion of Emer­gence (L-CAPE) pro­ject that started in late 2020 are (1) to apply data-an­a­lytic meth­ods to im­prove the in­for­ma­tion that is avail­able to op­er­a­tors in the con­trol room, and (2) to use ma­chine learn­ing to au­to­mate the la­bel­ing of out­age types as they occur and dis­cover pat­terns in the data that could lead to the pre­dic­tion of out­ages. We pre­sent an overview of the chal­lenges in deal­ing with time-se­ries data from 2000+ de­vices, our ap­proach to de­vel­op­ing an ML-based au­to­mated out­age la­bel­ing sys­tem, and the sta­tus of aug­ment­ing op­er­a­tions by iden­ti­fy­ing the most likely de­vices pre­dict­ing an out­age.  
poster icon Poster WEPA40 [1.870 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-WEPA40  
About • Received ※ 03 August 2022 — Revised ※ 12 August 2022 — Accepted ※ 17 August 2022 — Issue date ※ 31 August 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
THXD2 6D Phase Space Diagnostics Based on Adaptive Tuning of the Latent Space of Encoder-Decoder Convolutional Neural Networks controls, solenoid, feedback, electron 837
 
  • A. Scheinker
    LANL, Los Alamos, New Mexico, USA
 
  We pre­sent a gen­eral ap­proach to 6D phase space di­ag­nos­tics for charged par­ti­cle beams based on adap­tively tun­ing the low-di­men­sional la­tent space of gen­er­a­tive en­coder-de­coder con­vo­lu­tional neural net­works (CNN). Our ap­proach first trains the CNN based on su­per­vised learn­ing to learn the cor­re­la­tions and physics con­strains within a given ac­cel­er­a­tor sys­tem. The input of the CNN is a high di­men­sional col­lec­tion of 2D phase space pro­jec­tions of the beam at the ac­cel­er­a­tor en­trance to­gether with a vec­tor of ac­cel­er­a­tor pa­ra­me­ters such as mag­net and RF set­tings. The in­puts are squeezed down to a low-di­men­sional la­tent space from which we gen­er­ate the out­put in the form of pro­jec­tions of the beam’s 6D phase space at var­i­ous ac­cel­er­a­tor lo­ca­tions. After train­ing the CNN is ap­plied in an un­su­per­vised adap­tive man­ner by com­par­ing a sub­set of the out­put pre­dic­tions to avail­able mea­sure­ments with the error guid­ing feed­back di­rectly in the low-di­men­sional la­tent space. We show that our ap­proach is ro­bust to un­seen time-vari­a­tion of the input beam and ac­cel­er­a­tor pa­ra­me­ters and a study of the ro­bust­ness of the method to go be­yond the span of the train­ing data.  
slides icon Slides THXD2 [19.086 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-THXD2  
About • Received ※ 18 July 2022 — Revised ※ 05 August 2022 — Accepted ※ 08 August 2022 — Issue date ※ 09 August 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
FRXE1 Bayesian Algorithms for Practical Accelerator Control and Adaptive Machine Learning for Time-Varying Systems controls, feedback, experiment, electron 921
 
  • A. Scheinker
    LANL, Los Alamos, New Mexico, USA
  • R.J. Roussel
    SLAC, Menlo Park, California, USA
 
  Par­ti­cle ac­cel­er­a­tors are com­pli­cated ma­chines with thou­sands of cou­pled time vary­ing com­po­nents. The elec­tro­mag­netic fields of ac­cel­er­a­tor de­vices such as mag­nets and RF cav­i­ties drift and are un­cer­tain due to ex­ter­nal dis­tur­bances, vi­bra­tions, tem­per­a­ture changes, and hys­tere­sis. Ac­cel­er­ated charged par­ti­cle beams are com­plex ob­jects with 6D phase space dy­nam­ics gov­erned by col­lec­tive ef­fects such as space charge forces, co­her­ent syn­chro­tron ra­di­a­tion, and whose ini­tial phase space dis­tri­b­u­tions change in un­ex­pected and dif­fi­cult to mea­sure ways. This two-part tu­to­r­ial pre­sents re­cent de­vel­op­ments in Bayesian meth­ods and adap­tive ma­chine learn­ing (ML) tech­niques for ac­cel­er­a­tors. Part 1: We in­tro­duce Bayesian con­trol al­go­rithms, and we de­scribe how these al­go­rithms can be cus­tomized to solve prac­ti­cal ac­cel­er­a­tor spe­cific prob­lems, in­clud­ing on­line char­ac­ter­i­za­tion and op­ti­miza­tion. Part 2: We give an overview of adap­tive ML (AML) com­bin­ing adap­tive model-in­de­pen­dent feed­back within physics-in­formed ML ar­chi­tec­tures to make ML tools ro­bust to time-vari­a­tion (dis­tri­b­u­tion shift) and to en­able their use fur­ther be­yond the span of the train­ing data with­out re­ly­ing on re-train­ing.  
slides icon Slides FRXE1 [34.283 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-NAPAC2022-FRXE1  
About • Received ※ 08 August 2022 — Revised ※ 10 August 2022 — Accepted ※ 12 August 2022 — Issue date ※ 27 September 2022
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)