Testing the data framework for an AI algorithm in preparation for high data rate X-ray facilities Hongwei Chen124 Sathya R. Chitturi123 Rajan Plumley125 Lingjia Shen12 Nathan C. Drucker126

2025-05-02 0 0 1.38MB 9 页 10玖币

侵权投诉

Testing the data framework for an AI algorithm in

preparation for high data rate X-ray facilities

Hongwei Chen1,2,4∗, Sathya R. Chitturi1,2,3∗, Rajan Plumley1,2,5∗, Lingjia Shen1,2, Nathan C. Drucker1,2,6,

Nicolas Burdet1,2, Cheng Peng2, Sougata Mardanya7, Daniel Ratner1, Aashwin Mishra1, Chun Hong Yoon1,

Sanghoon Song1, Matthieu Chollet1, Gilberto Fabbris8, Mike Dunne1, Silke Nelson1, Mingda Li9,

Aaron Lindenberg2,3, Chunjing Jia2, Youssef Nashed1, Arun Bansil4, Sugata Chowdhury7,

Adrian E. Feiguin4, Joshua J. Turner1,2, Jana B. Thayer1

1Linac Coherent Light Source, SLAC National Accelerator Laboratory, Menlo Park, USA

2Stanford Institute for Materials and Energy Sciences, Stanford University, Stanford, USA

3Department of Materials Science and Engineering, Stanford University, Stanford, USA

4Department of Physics, Northeastern University, Boston, USA

5Department of Physics, Carnegie Mellon University, Pittsburgh, USA

6School of Engineering and Applied Sciences, Harvard University, Cambridge, USA

7Department of Physics and Astrophysics, Howard University, Washington, USA

8Advanced Photon Source, Argonne National Laboratory, Argonne, USA

9Department of Nuclear Science & Engineering, The Massachusetts Institute of Technology, Cambridge, USA

Abstract—The advent of next-generation X-ray free electron

lasers will be capable of delivering X-rays at a repetition rate ap-

proaching 1 MHz continuously. This will require the development

of data systems to handle experiments at these type of facilities,

especially for high throughput applications, such as femtosecond

X-ray crystallography and X-ray photon ﬂuctuation spectroscopy.

Here, we demonstrate a framework which captures single shot

X-ray data at the LCLS and implements a machine-learning

algorithm to automatically extract the contrast parameter from

the collected data. We measure the time required to return the

results and assess the feasibility of using this framework at high

data volume. We use this experiment to determine the feasibility

of solutions for ‘live’ data analysis at the MHz repetition rate.

Index Terms—LCLS-II, X-ray, Machine Learning, Experimen-

tal Design, High Performance Computing

I. INTRODUCTION

X-ray Free Electron Laser (XFEL) light sources are scien-

tiﬁc user facilities with the goal of investigating atomic scale

processes at ultrafast, femtosecond (1 femtosecond = 10−15

seconds) time-scales. [1]–[6]. In a typical XFEL experiment,

researchers from around the world are given access to the

facility for a limited amount of beamtime to perform experi-

ments at one of its many specialized scientiﬁc instrument end-

stations. Awarded beamtime is highly sought-after, and it is

not unusual for users to wait months or years to carry out

their experiments. The experiments are often highly complex,

involving a number of additional capabilities, such as advanced

laser systems, cryogenics, precision motion, ultra-high vac-

uum, high magnetic or electric ﬁelds, computational support

through both controls and analysis, and highly sensitive sam-

ples. Furthermore, the measurements involving the incident X-

ray laser radiation must be remotely controlled from outside

*Authors Contributed Equally

the local laboratory setup for the safety of the experimentalists.

With a period of beamtime access typically lasting between

12-60 hours, it is crucial that tools are in place so that XFEL

users are able to use their time efﬁciently to collect data and

perform the experiments.

Since their inception, XFELs have attracted much attention

due to the new ﬁelds of science which they enable based on

their high pulse intensity, short pulse duration, available X-ray

energies, and coherent properties [7]–[11]. A new generation

of these lasers is currently being operated, constructed or

planned at several sites around the world. With the increase

in repetition rate, entirely new experimental methods will

be made possible. Furthermore, increasing the XFEL repe-

tition rate will also address feasibility challenges for certain

experiments which currently take heroic efforts to perform,

such as photoemission spectroscopy [12], transient grating

spectroscopy [13], and resonant inelastic X-ray scattering [14].

One such technique is X-ray Photon Fluctuation Spec-

troscopy (XPFS), a method whereby one is able to use

temporally separated X-ray pulses to measure ﬂuctuations of

a system by probing changes at different ultrafast timescales

[15]. Although XPFS can provide an unprecedented level of

information in condensed matter systems, the data rates for

a typical experiment at the next-generation machines will be

intractable for realistic fast feedback during an experiment.

This hindrance is due to the fact that ∼106images on multiple

mega-pixel detectors will be collected per second and analyzed

‘on the ﬂy’, where the analysis must keep up with the input

data rate to extract the full value from the experimental time.

This volume of data is a dramatic increase over the current

XFEL data rate of hundreds or thousands of images per

second.

In this paper, we address this problem of the data rate posed

arXiv:2210.10137v1 [physics.data-an] 18 Oct 2022

by an increase of XFEL repetition rates by using the LCLS-I

data pipeline to run an experiment on a prototypical XPFS

experiment and assess the scalability of our analysis methods.

We measure scattered photons forming a speckle pattern

from a Van der Waals antiferromagnet [16]. The scattered

photons are measured through the Data Acquisition (DAQ)

system at the LCLS, reduced using a suite of tools called

small-data tools, and analyzed using a new machine-learning

algorithm running on a GPU system [17]. We benchmark

the performance of the algorithm running within the LCLS-

I analysis pipeline to evaluate its suitability for online data

reduction and to decipher potential barriers for efﬁcient use of

the data systems at the data rates of the emerging generation

of XFEL facilities such as LCLS-II, the European XFEL, and

the SHINE facility — up to the maximum rate of a continuous

MHz data stream. We use these results to project the needs

for capturing similar experimental data in the near future, and

we anticipate how to include Bayesian optimization in this

process to inform real-time steering of the experiment based

on comparisons to theory and prior data. These results will be

important for on-the-ﬂy decision making and eventually, for

experimental control at sophisticated XFEL beamlines.

II. MOTIVATION

There are many applications where the method described

here can be implemented. In general, these applications rely

on the unique properties of XFELs but may also be extensible

to the larger number of 4th generation synchrotron facilities

as well. We focus here on the speciﬁc application of XPFS,

where the temporal characteristics of the method match the

natural timescale of thermal or quantum ﬂuctuations of the

different degrees of freedom in the system being studied.

This has been shown by the progress made by XPFS in the

ﬁeld of topological magnets, but has also been applied to

other systems of interest as well, such as the behavior of

soft, disordered matter like colloidal nanoparticles [18], and

the exotic properties of water [19]. In a series of studies

[20]–[23], it was demonstrated that magnetic skyrmions –

spin textures with important topological properties for fu-

ture electronic devices [24] – ﬂuctuate in a distinct way on

nanosecond timescales. Moreover, these ﬂuctuations bare close

similarity with the exotic physics observed in other seemingly

disconnected systems, e.g. high-temperature superconductors

[22]. The discovery of these hidden connections highlights

the power of XPFS, and more importantly, urgently calls for

its implementation at the next-generation high-repetition rate

XFEL facilities.

Besides topological magnetism, many other areas within

the condensed matter community are ripe for the implemen-

tation of XPFS, such as the potential for addressing quantum

criticality. Unlike classical thermodynamical phase transitions,

a quantum phase transition (QPT) occurs at absolute zero

and is triggered by a non-thermal parameter [25]. It is solely

governed by Heisenberg’s uncertainty principle, i.e. quantum

ﬂuctuations. This novel type of physics is thought to be

important for unconventional superconductivity [26]. However,

the exact role of quantum ﬂuctuations in materials of this

class remains highly debated. One advantage of XPFS is that

it can offer a direct evaluation of quantum order parameter

ﬂuctuations covering a broad time window ranging from the

femtosecond level to hundreds of nanoseconds. The study

of the dynamics of the order parameter, for instance, could

be explored far away from to the regions in close vicinity

to the quantum critical point. This capability of accessing

ﬂuctuations on these distinct timescales and over these ranges

is not captured by any other current X-ray techniques.

III. APPLICATION: XPFS

XPFS relies on carefully comparing sequential scattering

patterns of outgoing X-ray photons from the sample after it is

illuminated by the incident XFEL beam. This process is akin

to taking a series of photographs in very rapid succession,

and adding them together while studying the “fuzziness” of

the image, to extract out valuable dynamics occurring in the

system. The primary quantitative metric used for this compar-

ison is the contrast of the scattering intensity pattern collected

by the X-ray detector. Since the resultant scattering intensity

or “speckle” pattern from the sample is directly related to its

microscopic structure, one is able to draw conclusions about

dynamical changes in the sample by monitoring the change in

contrast [27], [28]. This method of analyzing changes in the

sample is incredibly powerful for ultrafast physics experiments

because its temporal sensitivity is only limited by the X-ray

pulse separation time and not the readout time of the detector

[29].

However, accurate contrast determination requires a large

number of measurements as well as a spatially precise method

for mapping the photon hits in the detector. This becomes

especially important in the so-called photon-hungry regime,

where the next-generation XFEL facilities beneﬁt strongly.

The challenge associated with mapping the single photon

distribution is due to the problem of X-ray photons exciting

a charge cloud in the detector sensor that is comparable to

or larger than a single pixel [15]. This “bleeding” of charge

into adjacent pixels creates characteristic artifacts where the

signal associated with a single photon is spread across multiple

pixels, creating a degree of uncertainty in the scattering

patterns. This is dealt with using a greedy guess algorithm for

reconstructing the photons from their charge clouds [30], [31].

The statistical aspect of the data collection can be understood,

in the sparse photon limit, by the following equation, where the

standard deviation σcfor contrast C0scales with the average

photons/speckle ¯

k, number of speckles Nspeckle, and number

of image frames collected Nframe as [31]:

σc

C0¯

k2(1 + C0)

NspeckleNframe 1/2

(1)

Both criteria for statistical volume and accurately

reconstructed photon-placement must be satisﬁed in order for

contrast to be reliably determined. In the case of an XFEL

experiment where users are relying on analysis feedback in

order to make critical and timely decisions related to what

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

TestingthedataframeworkforanAIalgorithminpreparationforhighdatarateX-rayfacilitiesHongweiChen1;2;4,SathyaR.Chitturi1;2;3,RajanPlumley1;2;5,LingjiaShen1;2,NathanC.Drucker1;2;6,NicolasBurdet1;2,ChengPeng2,SougataMardanya7,DanielRatner1,AashwinMishra1,ChunHongYoon1,SanghoonSong1,MatthieuChollet1,Gil...

展开>> 收起<<

Testing the data framework for an AI algorithm in preparation for high data rate X-ray facilities Hongwei Chen124 Sathya R. Chitturi123 Rajan Plumley125 Lingjia Shen12 Nathan C. Drucker126.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Testing the data framework for an AI algorithm in preparation for high data rate X-ray facilities Hongwei Chen124 Sathya R. Chitturi123 Rajan Plumley125 Lingjia Shen12 Nathan C. Drucker126

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: