Monitoring MBE substrate deoxidation via RHEED image-sequence analysis by deep learning Abdourahman Khaireh-Walieh1Alexandre Arnoult1Sébastien Plissard1and Peter R. Wiecha1 1LAAS-CNRS Université de Toulouse CNRS UPS F-31400 Toulouse France

2025-05-02 0 0 968.49KB 8 页 10玖币

侵权投诉

Monitoring MBE substrate deoxidation via RHEED image-sequence analysis by deep learning

Abdourahman Khaireh-Walieh,1Alexandre Arnoult,1Sébastien Plissard,1and Peter R. Wiecha1, ∗

1LAAS-CNRS, Université de Toulouse, CNRS, UPS, F-31400 Toulouse, France

Reﬂection high-energy electron diffraction (RHEED) is a powerful tool in molecular beam epitaxy (MBE),

but RHEED images are often difﬁcult to interpret, requiring experienced operators. We present an approach

for automated surveillance of GaAs substrate deoxidation in MBE reactors using deep learning based RHEED

image-sequence classiﬁcation. Our approach consists of an non-supervised auto-encoder (AE) for feature ex-

traction, combined with a supervised convolutional classiﬁer network. We demonstrate that our lightweight

network model can accurately identify the exact deoxidation moment. Furthermore we show that the approach

is very robust and allows accurate deoxidation detection during months without requiring re-training. The main

advantage of the approach is that it can be applied to raw RHEED images without requiring further information

such as the rotation angle, temperature, etc.

Keywords: RHEED, deep learning, molecular beam epitaxy, substrate deoxidation

I. INTRODUCTION

Reﬂection high-energy electron diffraction (RHEED) is a

widely used in-situ control method in molecular beam epi-

taxy (MBE) [1–4]. RHEED diffraction patterns provide in-

formation about the crystal surface with atomic resolution

and as the ultra high vacuum in typical growth chambers al-

lows an easy integration of electron beam systems in MBEs,

RHEED has become a standard in-situ characterization instru-

ment in MBE, enabling unprecedented accuracy in monitoring

the crystal growth. RHEED is highly sensitive to several key

MBE parameters like the growth rate, the crystal structure, the

lattice parameter and strain, etc [5–9]. However, RHEED im-

ages can be difﬁcult to interpret since the diffraction patterns

produce information in the Fourier-space. Furthermore, the

actual recorded patterns are very sensitive to calibration, and

often also dynamic variations in the patterns over several time-

scales contain valuable information, rendering their analysis

even more challenging. Real-time exploitation of RHEED

data is therefore often limited to easily accessible informa-

tion like the deposition rate. Sophisticated analysis is usually

done a posteriori, on recorded RHEED images or videos. Due

to the complexity of the task, RHEED interpretation usually

requires experienced operators, possessing years of machine-

speciﬁc training.

A common application of RHEED is the monitoring of the

native oxide removal from commercial substrates prior crys-

tal growth. Surface oxidation of a few nanometers due to ex-

posure to oxygen is unavoidable during transport of epitaxial

substrates, which renders their surface non-crystalline. This

oxide needs to be removed before any epitaxial material de-

position, which is usually done by heating. In the case of gal-

lium arsenide (GaAs), the substrate is slowly heated to around

610◦C, while stabilizing the crystal with a constant arsenic

ﬂux of around 1.2×10−5Torr, to avoid As evaporation [10].

Once the oxide is removed, in order to avoid damaging of

the crystal, further temperature ramping needs to be stopped,

usually temperature is in fact decreased. To detect the deoxi-

dation, the MBE operator supervises the RHEED image dur-

∗e-mail : pwiecha@laas.fr

ing temperature increase, and once the diffraction pattern of a

crystalline surface starts to form, the operator manually ends

the heating procedure. Not only is the constant presence of the

operator required, due to its manual character the deoxidation

procedure is furthermore error-prone. Automatic detection of

the deoxidation is challenging, ﬁrst because RHEED patterns

are often weak since the raw substrate surfaces are not atom-

ically ﬂat, and second because the RHEED image contrast is

dependent on some parameters like ﬁlament current or elec-

tron beam angle, and hence is not exactly constant in each

run. Finally, the substrate is usually laying on a rotating sam-

ple holder, hence the RHEED pattern constantly changes.

Methods from artiﬁcial intelligence including deep learning

are increasingly applied to nano- and material-science [11–

15]. Recently, ﬁrst attempts have been reported to use sta-

tistical methods and machine learning for RHEED image in-

terpretation [16–20]. Inspired by these pioneering works, we

propose to use a deep learning (DL) approach for classiﬁca-

tion of oxidized and deoxidized substrates via their RHEED

patterns, to resolve the above described problems. As men-

tioned above, due to the sample rotation, the RHEED signal

can conﬁdently indicate deoxidation only during short mo-

ments, when the electron beam is aligned with a lattice di-

rection of the crystal. In contrast to recent propositions to

use DL with RHEED for surface reconstruction identiﬁcation

[19,20], we therefore propose a model which analyzes se-

quences of RHEED patterns (i.e. videos), instead of single

images. To this end, we propose a two-stage deep learning

model: The ﬁrst stage is an autoencoder, which compresses

each full-resolution RHEED image into a low-dimensional la-

tent vector. The second stage subsequently determines the

oxidation state for a sequence of such latent vectors, hence

for the compressed representation of a short RHEED video

sequence. We provide a detailed analysis of the required la-

tent and sequence lengths and demonstrate the accuracy of

the model as well as its stability over a period of more than 6

months between training data sampling and testing. Finally,

we provide online the data, codes as well as pre-trained mod-

els to reproduce our results [21].

arXiv:2210.03430v2 [cond-mat.mes-hall] 15 Dec 2022

sample RHEED

screen

RHEED video

electron beam

deep learning

model

surface:

oxidized ?

deoxidized ?

(a) (b)

oxidized

deoxidized

encoder decoder

latent

(dim. N)

RHEED

image reconstr.

image

...

classiﬁersequence

L latents

(d) RHEED image example sequence - deoxidized GaAs surface

Figure 1. Deoxidation detection problem. (a) Short sequences of the RHEED video obtained from a rotating GaAs substrate are analyzed by

a deep learning neural network model (NN), to determine if the substrate deoxidation process has terminated. (b) The DL model consists of

two stages. An autoencoder encodes every single RHEED image into a latent vector of dimension Nwith the goal of reconstructing the original

image. The latent thus contains all relevant features of the RHEED image. Sequences of Llatent vectors are then used for classiﬁcation of

deoxidized surfaces. The classiﬁer network thus analyses short RHEED videos covering a certain rotation angle. (c) Example sequence of

10 consecutive raw RHEED images obtained from an oxidized GaAs surface. (d) Example sequence of 10 consecutive raw RHEED images

obtained after deoxidation from the same GaAs surface as shown in (c).

II. RESULTS AND DISCUSSION

A. Substrate deoxidation classiﬁcation Problem

On commercial substrates, a native oxide layer of a few

nanometers encapsulates the crystal surface. The RHEED

electron beam does not penetrate through this oxide, and

hence does not reach the crystalline lattice (of in our case

GaAs). The electrons are thus not diffracted, but scattered.

Independent of the sample rotation angle, the RHEED signal

on the ﬂuorescent screen is diffuse and no diffraction pattern

occurs (c.f. Fig. 1c). Without oxide layer on the other hand,

the RHEED electrons are diffracted by the atomic lattice of

the now crystalline surface. However, the transition from oxi-

dized to deoxidized is not instantaneous and during the deox-

idation the classiﬁcation is often difﬁcult. Furthermore, due

to the rotation of the sample, the diffraction pattern continu-

ously changes and, especially during the deoxidation process,

the pattern arises not similarly clearly for different rotation an-

gles. Our operator classiﬁes the surface as deoxidized when a

clear diffraction pattern occurs repeatedly during at least one

full rotation cycle of the substrate.

The general problem is schematically depicted in Figure 1a.

Our goal is to precisely determine the moment of full oxyde

removal from a GaAs substrate by monitoring the RHEED

pattern during the deoxidation process.

However, as mentioned above, the image dynamics due to

the constant rotation of the sample is a challenge for an algo-

rithmic evaluation. Furthermore, disordered bright spots can

occur also from oxidized surfaces (see Fig. 1c). Therefore,

an algorithmic classiﬁcation is not entirely trivial. By feeding

short video sequences of several consecutive RHEED images

to a classiﬁcation neural network, we aim at determining the

oxidation state of the substrate surface, in order to reduce the

necessity of human supervision of the substrate cleaning pro-

cess.

B. Dataset

To train a neural network on deoxidation reconnaissance,

we generate a training dataset by capturing RHEED videos

before and after the oxide removal procedure. The images

are collected in real time at 24 frames per second, while the

sample rotates with 12 rounds per minute, hence we capture

120 images per full rotation. The RHEED video is thereby

captured image by image, using a CMOS Camera (Allied

Vision Manta G319B) with 4 ×4 pixel binning, resulting in

raw images of 416 ×444 pixels at 12 bit grayscale intensity

resolution. Those images are simply converted to 8 bit for-

mat and scaled to 100 ×100 pixels. In total we collected

videos containing a total of 7644 RHEED images from ﬁve

substrate oxide removal procedures within a period of a few

days. 3110 of these images correspond to deoxidized sur-

faces, the rest are images from GaAs surfaces which were

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：

MonitoringMBEsubstratedeoxidationviaRHEEDimage-sequenceanalysisbydeeplearningAbdourahmanKhaireh-Walieh,1AlexandreArnoult,1SébastienPlissard,1andPeterR.Wiecha1,1LAAS-CNRS,UniversitédeToulouse,CNRS,UPS,F-31400Toulouse,FranceReectionhigh-energyelectrondiffraction(RHEED)isapowerfultoolinmolecularbeame...

展开>> 收起<<

Monitoring MBE substrate deoxidation via RHEED image-sequence analysis by deep learning Abdourahman Khaireh-Walieh1Alexandre Arnoult1Sébastien Plissard1and Peter R. Wiecha1 1LAAS-CNRS Université de Toulouse CNRS UPS F-31400 Toulouse France.pdf

共8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！

Monitoring MBE substrate deoxidation via RHEED image-sequence analysis by deep learning Abdourahman Khaireh-Walieh1Alexandre Arnoult1Sébastien Plissard1and Peter R. Wiecha1 1LAAS-CNRS Université de Toulouse CNRS UPS F-31400 Toulouse France

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: