Remote multi-user control of the production of Bose-Einstein condensates for research and education
J S Laustsen, R Heck, O Elíasson, J J Arlt, J F Sherson, C A Weidner
RRemote multi-user control of the production ofBose-Einstein condensates for research and education
J S Laustsen, R Heck, O Elíasson, J J Arlt, J F Sherson and C A Weidner
Department of Physics and Astronomy, Aarhus University, 8000 Aarhus C, DenmarkE-mail: [email protected]
28 January 2021
Abstract.
Remote control of experimental systems allows for improved collaborationbetween research groups as well as unique remote educational opportunities accessible bystudents and citizen scientists. Here, we describe an experiment for the production andinvestigation of ultracold quantum gases capable of asynchronous remote control by multipleremote users. This is enabled by a queuing system coupled to an interface that can be modifiedto suit the user, e.g. a gamified interface for use by the general public or a scripted interfacefor an expert. To demonstrate this, the laboratory was opened to remote experts and thegeneral public. During the available time, remote users were given the task of optimising theproduction of a Bose-Einstein condensate (BEC). This work thus provides a stepping stonetowards the exploration and realisation of more advanced physical models by remote experts,students and citizen scientists alike.
Keywords : Remote experiment control, ultracold atoms, BEC
1. Introduction
Ultracold quantum gases have become one of the prime platforms for simulatingtechnologically relevant quantum systems within the last decades. In particular, extremelyclean and pure quantum model systems can be realised that offer a high degree ofcontrollability with respect to parameters such as the atoms’ interaction strength andtemperature. This progress has led to increasingly poweful and complex experiments inlattice-based quantum simulation [1, 2], the simulation of strongly-correlated condensedmatter systems [3], and quantum computing with Rydberg atoms [4], among others, renderingcold atoms a fantastic platform for the development of technologies that will drive the secondquantum revolution [5].Collaboration between experimental and theoretical groups is an essential part ofdeveloping and evaluating models applicable to quantum simulation experiments. To optimiseexperimental procedures, it may indeed be beneficial to use dedicated, automated protocolsdeveloped by theory groups. Opening the laboratory to direct remote control by collaboratorsmay thus increase the efficiency of such collaborative efforts. Moreover, a remote control a r X i v : . [ c ond - m a t . qu a n t - g a s ] J a n system opens up new possibilities regarding outreach to students and the general public. Byallowing a broad audience of non-expert users to control some experimental parameters, onecan imagine a number of scenarios geared towards public outreach and education. First, thepublic can take part in citizen science experiments, and, in particular, previous work using thesystem described here shows that valuable insight into cognitive science can be gained [6].Secondly such platforms can be used to educate and engage non-experts in quantum physics,e.g. by allowing students access to cutting-edge research laboratories regardless of wherethe laboratory is physically located. In both cases this creates the need for an intuitive userinterface which allows users to focus on the essential parts and hides the technical details.At the same time, the experimental system must also contain the infrastructure to handle theinput from one of many users and return the relevant results to the correct user. A number ofthese open platforms already exist, including the IBM Quantum Experience [7], and the openavailability of this platform has allowed for the production of a number of research articles(see, e.g., Refs. [8, 9, 10, 11, 12]), educational material [13], and games [14].In principle, any experimental control program can be modified for remote control viathe addition of a remote server and a suitable front-end for the user. In terms of experimentalcontrol programs, several publicly available systems for cold atom experiments have beenpublished [15, 16, 17, 18, 19]. In addition, numerous commercial options are available,such as systems by ColdQuanta [20], MLab [21], ADwin [22] and IQ Technologies [23]that can be purchased together with suitable hardware. All of these control systems havesub-microsecond control of digital and analog channels and some allow for direct outputof radio frequency (RF) signals. Additionally, they typically allow for communication withexternal hardware through different protocols or via implementation of appropriate drivers.These criteria define the typical minimum viable product for useful cold-atom experimentcontrol. Software for camera control and analysis of the images enables some systems tooptimise experimental performance in a closed loop optimisation of experimental parameters.Moreover, all of these systems are remotely-controllable either directly or via simple screen-sharing protocols. However, to our knowledge, none of these control programs had been usedin a multi-user setting where several users simultaneously remotely controlled an experimentthrough the use of the aforementioned server and front-end, with the exception that, whilepreparing this manuscript, we became aware of the Albert system built by ColdQuanta thatcame online in late 2020 [24]. This work thus represents the growing commercial andacademic interest in remote control of cold atom systems.Here we discuss the implementation of a remote controlled experiment usable by singleexpert user or multiple non-expert users accessing the experiment. Previously, we havedocumented the challenges that we provided to our users, as well as the main findings thatarose from this work [6]. However, we have not explained the underlying system architectureand the overarching possibilities that this gives rise to in research and education. The generalknowledge of these details is crucial for other groups to implement similar systems, and thisis what we focus on in this work. In both of the use-cases considered here, there is a needfor a queuing system for the input sequences and the return of the results. When consideringmulti-user access there is also the need to track the sender throughout the process of queuing,performing the experiment, analysis and reporting the results. The infrastructure of theexperiment also allows for multiple expert users and this option will be explored in futurework. For instance, one could imagine running multiple collaborative efforts simultaneously.This paper is organised as follows: In the first section the software enablingremote control is presented. Following this, the experimental sequence and its technicalimplementation is described. We then describe the two different implementations of remoteuser access that were used in previous work: single- and multi-user control [6]. The lastsection concludes and provides an outlook.
2. The control software
The experimental control system is LabVIEW-based and capable of being expanded as newhardware is added to the experiment. A field-programmable gate array (FPGA, PCI-7813R)is used to control 70 digital and 48 analog channels through 4 digital to analog convertermodules (NI 9264, NI 9263). In addition, the system can communicate with hardware driversto other hardware such as motion stages, piezoelectric devices and RF synthesisers. Thus, ourcontrol system meets the aforementioned criteria for usability in a cold atom experiment.The control program is based on a server/client architecture. The server controls allhardware, including the FPGA, and the client provides an interface for the user and compilesthe programmed sequence. On the client side the sequence is built of waves which correspondto the output of a given digital or analog channel, a GPIB command or a command througha hardware driver. Regularly-used sequences of waves can be collected and collapsed into blocks , e.g., the commands required to load atoms into a trap or image an atom cloud. For eachwave and block, externally-accessible variables can be declared, e.g. the frequencies of theRF tones applied to acousto-optic modulators (AOMs) or the duration of the RF pulse appliedto the AOM. This allows the user to create sequences with an adaptive level of abstraction.For instance one can hide the exact technical implementation of experimental steps in a blockbut keep the essential control parameters accessible, which is useful for reducing the cognitiveload of a remote user.An example of a block used for absorption imaging of ultracold atoms is shown in Fig. 1,where smaller blocks are incorporated. The waves and blocks are ordered in a tree structurethat controls the timing of an experiment. The tasks are performed from top to bottom insuch a tree. Any waves or blocks on indented branches are performed simultaneously, anddelays can be defined within individual elements for more precise control of relative timing.Initialised outputs may be defined such that they either hold their last value or are reset to adefault value after a given time. Thus the user need only handle the values of relevant outputsat any given point. Wave and block variables can be scanned individually or jointly in single-and multi-dimensional scans, respectively. Loops are also available where a subset of blocksis repeated while one or several variables are changed. For example, the user can loop thecapture of atoms in a trap while changing a given parameter value during each loop iteration,effectively performing a parameter scan within a single realisation of the experiment.The novel aspects of the control system lie in its capability for communication withremote users. This includes loading sequences from a queue either created by a single user ormultiple different users. After a remotely-requested sequence is performed, relevant results(e.g. atom number) are sent back to the user who designed the sequence. To make the remotecontrol as flexible as possible, the control software does not provide any user interface forthe remote user but communicates with stand-alone interfaces. Thus a remote user can easilyset up closed-loop optimisation by linking the returned results into a script running a givenoptimisation algorithm that then generates the next desired sequence, as described in detailbelow.
3. The experiment
To demonstrate the use of the control system and the communication necessary for multi-user operation, we conducted an experiment in which remote users create a Bose-EinsteinCondensate (BEC). The experimental system is described in Refs. [25, 26] and only its mainfeatures are described here.The experimental sequence starts by precooling a cloud of Rb-87 atoms in a 3D MOT.Here the atoms are laser-cooled and trapped via a combination of light pressure and magneticfield gradients. Subsequently polarisation gradient cooling is performed and the atoms areoptically pumped to the low-field-seeking | F , m F (cid:105) = | , (cid:105) state. The atoms are then trappedin a magnetic quadrupole trap generated by a pair of coils in an anti-Helmholtz configuration.These coils are mounted on a rail and are used to transport the atoms through a differentialpumping stage to the final chamber. Here the atoms are evaporatively cooled by transferringthe hottest atoms to a high-field-seeking sublevel. By the end of the evaporation sequencethe atoms have a temperature of roughly 30 µK. Subsequently, a crossed optical dipole trap(CDT) consisting of two laser beams (wavelength λ = / e waists of 45 µm and85 µm) is superimposed on the atoms. After the final evaporation stage, the atom cloud isreleased from the trap and an absorption image is recorded after a TOF. If the user-definedevaporation sequence is effective, the cloud is condensed and a BEC is visible in the image.For the remote experiments reported here, the control parameters available to the usersare the laser powers of both laser beams forming the CDT and the current in the quadrupolecoils as a function of time. This configuration allows the user to cool the atoms using forcedevaporative cooling either in a pure CDT [27], in a so-called hybrid trap consisting of thequadrupole magnetic field and one of the dipole beams [28], or any combination of the two.The depth and geometry of the trap depends on these parameters in a non-trivial optimal way,providing an opportunity for external optimisation, the goal of which is to produce the largestpossible BEC. For both expert and non-expert users a limitation of the available control spaceis necessary as only a small fraction of the full control landscape will yield a BEC.
4. Two cases of remote user control
For a remotely-controllable system to be useful, appropriate user interfaces must bedeveloped, and each user class has different requirements to optimally facilitate the
Imaging block Delay (ms) Type A/DOpen imaging shutterSet img. AOM frequencyTurn off CDT AOMTake image with light + atomsTake image with lightClose imaging shutterTrigger camera (dark image) WWWBBWWTOF300300 DADDD Take image block Delay (ms) TypeTrigger cameraTurn on imaging AOM WW DDA/D
Figure 1.
An example of a block used take an absorption image at the end of an experimentalsequence (cf. Sec. 3). The block contains individual analog (A) and digital (D) waves (W),as well as two embedded blocks (B) used to take the absorption and background images. Theblock runs from the top to bottom with indented elements running in parallel with the elementabove. In this block, the imaging shutter is opened while the frequency applied to the imagingAOM is set and the CDT AOM is turned off (thus turning off the CDT itself and dropping theatoms from the trap). Then, after a variable time-of-flight (TOF) the next block (blue squares,with zoom-in to the right of the main block) simultaneously pulses the imaging AOM andtriggers the camera shutter to take the absorption image of the atoms. After 300 ms of cameraprocessing time, the same block takes a background image without atoms. The imaging shutteris closed, and after an additional 300 ms, the camera is triggered again to take the dark imagewithout any light present. Subtracting the absorption image from the background and darkimages reveals the atom signal. interaction. For experts a scripted interface can be an advantage as complex algorithms canbe directly implemented. A more visual interface of the control (for instance in a game-like setting) is needed for non-expert users. Importantly, a different program structure isneeded when handling input from a single user or multiple users. In what follows, we describetwo different implementations of our remote control geared towards single expert users andasynchronous use among the general public, respectively. In this section, we elaborate oneach of these cases. Again, note that the data presented here is drawn from the same source asour initial work [6], and detailed research results can be found there. Here, we focus on moretechnical aspects of the experimental implementation and execution.
In our first implementation of remote control, an expert user optimised the evaporation usingthe so-called dressed chopped random-basis (dCRAB) optimisation algorithm [29]. Notethat the algorithm was implemented on the user side, so our implementation of remoteexpert control is algorithm agnostic. Here the user had access to the CDT laser powers andquadrupole coil currents as a function of time. Sequences of waveforms corresponding tothe parameter values were created as text files and sent to the experimental control programthrough a folder on a cloud drive and placed in a queue. Even for a single user, a queue isnecessary due to the relatively long (30 s) cycle time of the experiment. The queue operatedon the first-in-first-out (FIFO) principle, allowing the user to submit several sequencessimultaneously and easily keep track of the outputs; this is useful, e.g., when initialisingthe initial simplex for Nelder-Mead optimisation.For each user-accessible parameter, the parameter values can be defined at any desired
Figure 2.
A screenshot of the interface used in the
Alice Challenge , showing (a) the splineeditor used to create the ramps of the laser powers and coil current, shown on a logarithmicscale, (b) top score list, (c) latest executed sequences, and (d) the control buttons, includingthe estimated wait time until the submitted sequence is returned. time, while values between these times are linearly interpolated at the hardware level.Therefore the effective temporal resolution of the waveforms can be controlled by the user,and the total number of parameter/time pairs that can be used is ultimately limited by thememory of our FPGA.When a given sequence was ready to be run, the relevant experimental sequences weregenerated by reading in the waveforms the from text files generated by the expert user. Theexperiment was then run and the resulting image was analysed. From this image the BECatom number was extracted and returned to the expert user through the same cloud drive,again as a text file. This atom number was read in by the expert user and served as the costparameter closing the optimisation loop.
In the second implementation, called the
Alice Challenge , citizen scientists were given accessto the experiment. This subsection details the architectural considerations required for thechallenge as well as some statistics on user load in real time over the course of the challenge.This information is useful when considering the future implementation of similar systems.Citizen scientists were given access to the system via a gamified interface as shown inFig. 2, and this is used to provide more intuitive access to the parameter space. The interfacewas designed to visualise the ramps of the laser powers and coil currents sent by the citizenscientists to the experiment. The control values were normalised and presented for ease of
User AUser B
UIUI
ControlsystemQueue ImageAnalysisExptWebserver
Figure 3.
Schematic view of the data flow for the remote control of the experiment bymultiple, asynchronous outside users during the
Alice Challenge . Experimental sequences aresubmitted through a game-like user interface to a web server that subsequently sends themto the cloud-based queue in the order they are received (here, User A has submitted theirsequence first). Each submission has a unique User ID that is tracked throughout the process.The control system reads the oldest files via the FIFO principle and runs the correspondingexperiment. When image analysis is completed, the results are returned to the proper user viathe UI. use on a logarithmic axis in a spline editor where the user could manipulate the curves byclicking-and-dragging points along the curve. When the user was done editing the curves, thesequence was submitted and subsequently realised in the experiment.This was done in the following manner: The user sequence (encoded as a JSON filewith a unique user ID) was delivered to a web server. The web server then delivered thesequence to the cloud folder that served as the queue. When a sequence was ready to beevaluated, it was sent to the LabVIEW control system, where the JSON data was translatedinto waveform data identical to the type used in the single expert user configuration. This wasdone via a special optimisation class defined in LabVIEW that was responsible for extractingthe relevant parameters from the JSON file. Once a sequence was completed, the controlprogram wrote the results to another JSON file, inserted the relevant user ID, and stored it ina separate folder on the cloud. The webserver then delivered the results, and the backend ofthe game interface scaled the BEC atom number to a score which was displayed to the user.The score and corresponding sequence was also visible for other players who could then copythe sequence as inspiration when creating their own sequences.In contrast to the case of a single user, the web server was needed to track the run numberand user ID if multiple users were running the experiment simultaneously. A schematic viewof the multi-user data handling infrastructure used for the
Alice Challenge is presented in Fig.3. To ensure that the result of the experiment was linked to the right user sequence a check wasmade in the experimental control system such that the experimental sequence was repeated ifno result was returned for a given run.Moreover, the state of the experiment was checked by inserting an established benchmarksequence in the queue every tenth run. This benchmark sequence was known to create a BECunder stable experimental conditions. In the case of a problem, such as a laser failure, theexperiment was paused until the problem was solved. At the same time, the users wereinformed of the temporary delay caused by the disturbance. The benchmark sequence wasalso executed in case of an empty queue in order to keep the experiment in a stable condition. date r un r a t e [ / m i n ] date un i que u s e r s Figure 4.
The mean rate at which the experimental sequences were performed during the weekin which the experiment was open to non-expert outside users. The different plateaus arisedue to changes in the ramp time whereas the high run rates are an effect of synchronisationproblems. The green shading denotes the time period depicted in Fig. 5. In the inset theaccumulated number of unique citizen scientists that used the system throughout the week isshown. The date markers indicate midnight CET on a given day.
This also allows one to track overall experimental drifts, e.g. due to thermal fluctuations,which can be useful in advanced closed-loop optimisation schemes.The experiment was open to the public for a full week with only minor interruptions,resulting in total of 7577 submitted and evaluated sequences. Figure 4 shows the rate of theexperimental runs during this week. Over the course of the challenge, the preset duration ofthe evaporation ramps were varied. This allowed the citizen scientists to explore differentoptimisation landscapes, varying the challenge offered to them and keeping things interestingfor returning users.The different evaporation durations create some variation of the rate at whichexperiments were performed over the course of the week. In addition, when experimentalproblems caused the experiment to be paused, the rate decreased. It should also be noted thatthe peaks of high run rates were caused by synchronisation problems between the web serverand the control program. This problem was solved on the third day of the challenge, afterwhich none of the larger peaks are visible. The inset shows the progress of the accumulatednumber of unique users through out the week.Throughout the day the number of active users varied as players from several parts of theworld came online. Figure 5 shows the queue and number of active users on Friday evening(CET), one of the highest peaks in active users. Here we see that up to 15 unique users were time W a i t t i m e [ m i n ] un i que a c t i v e u s e r s Figure 5.
The queue time and number of active users during the busiest period of thechallenge. The blue trace shows the calculated waiting time from a submission of a sequenceuntil a result is given and the green trace shows the number of active users at any given point. active at any given time during the evening which created a wait time of above one hour. Asthe number of users declined, the length of the queue was slowly reduced. Since each usercould submit several different sequences at a time, the correlation between the number ofunique users and the queue length is nonlinear.Figure 6 shows a histogram of how many times a given BEC number was achieved. Wesee that most sequences submitted to the experiment result in the creation of a BEC. Thisis despite the fact that citizen scientists had limited insight to the physical system they werecontrolling to create the condensates.
5. Outlook
In future work, remote controlled optimisation of a system may be advantageous, asremote optimisation allows for easy implementation of advanced optimisation algorithms.Several programs are available that can implement closed-loop optimisation of cold-atomexperiments [15, 16, 17, 18, 19]. Students can also access such systems for educationalpurposes, as has already been done with quantum computers [13]. This allows students toexplore complex, cutting-edge research systems that are not accessible in many educationallearning laboratories.For remote users to be able to run optimise the experiment, the relevant experimentalcontrol parameters have to be easily controllable. Collaborative optimisation between several0 N BEC ( 10 ) O cc u r en c e Figure 6.
A histogram of how many times a given BEC atom number was obtained bythe sequences submitted by the users of the Alice challenge. Above 73% of the submittedsequences created a BEC. This data is also presented in Ref. [6]. remote users requires a structure that includes a multi-user queue and tracking the ID ofsubmitted sequences so that the results may be returned to the correct user. The controlprogram presented here can be expanded to give remote access to larger parts of the controlsequence or even the entire experiment. Thus, future work will give remote users expandedaccess, allowing them to tackle more advanced scientific problems in a research or educationalsetting. For example, with the new capabilities of the experiment to image single atomsusing a quantum gas microscope [30] in combination with spin addressing [31] and arbitrarylight potential generation techniques [32] the experiment can be used as an analog quantumsimulator with remote control capability.Such advanced control will require a more complex user interface since the number ofexperimental parameters would increase, rendering algorithmic optimisation more difficult.However other groups have shown that optimisation of such systems with large parameterspaces is possible using genetic algorithms [33, 34, 35] or machine learning methods such asneural networks [36], Gaussian processes [37] or evolutionary optimisation [38].
6. Acknowledgements
The authors would like to acknowledge Aske Thorsen for the development of the LabVIEWcontrol code used for the experiments described here.1
References [1] Gross C and Bloch I 2017
Science
Nature Reviews Physics Journal of Physics B: Atomic, Molecular and Optical Physics Journal of Physics B: Atomic, Molecular and Optical Physics PRX Quantum Proceedings of the National Academy of Sciences
E11231–E11237 ISSN 0027-8424, 1091-6490[7] IBM Quantum Experience https://quantum-computing.ibm.com/[8] Berta M, Wehner S and Wilde M M 2016
New Journal of Physics Physical Review Letters npj Quantum Information
46 ISSN 2056-6387[11] Wootton J R and Loss D 2018
Physical Review A IEEE Transactions on Computer-Aided Design of Integrated Circuitsand Systems arXiv:2012.09629 [physics,physics:quant-ph] ( Preprint )[14] Wootton J R 2017
Proceedings of the Second Gamification and Serious Games Symposium Review of Scientific Instruments Review of Scientific Instruments Review of Scientific Instruments Review of Scientific Instruments Preprint )[19] Perego E, Pomponio M, Detti A, Duca L, Sias C and Calosso C E 2018
Review of Scientific Instruments Journal of Physics B: Atomic, Molecular and Optical Physics Journal of Physics B: Atomic, Molecular and Optical Physics AdvancesIn Atomic, Molecular, and Optical Physics vol 42 (Elsevier) pp 95–170 ISBN 978-0-12-003842-8[28] Lin Y J, Perry A R, Compton R, Spielman I B and Porto J 2009
Physical Review A Physical Review A [30] Elíasson O, Laustsen J S, Heck R, Müller R, Weidner C A, Arlt J J and Sherson J F 2020 arXiv:1912.03079[cond-mat, physics:quant-ph] ( Preprint )[31] Weitenberg C, Endres M, Sherson J F, Cheneau M, Schauß P, Fukuhara T, Bloch I and Kuhr S 2011
Nature
Stochastic Optimization - Seeing the Optimalfor the Uncertain ed Dritsas I (InTech) ISBN 978-953-307-829-8[34] Lausch T, Hohmann M, Kindermann F, Mayer D, Schmidt F and Widera A 2016
Applied Physics B [35] Geisel I, Cordes K, Mahnke J, Jöllenbeck S, Ostermann J, Arlt J J, Ertmer W and Klempt C 2013
AppliedPhysics Letters
ISSN 00036951[36] Tranter A D, Slatyer H J, Hush M R, Leung A C, Everett J L, Paul K V, Vernaz-Gris P, Lam P K, BuchlerB C and Campbell G T 2018
Nature Communications Scientific Reports MachineLearning: Science and Technology1