MICRO LIB A LIBRARY OF 3D MICROSTRUCTURES GENERATED FROM 2D MICROGRAPHS USING SliceGAN Steve Kench1 Isaac Squires1 Amir Dahari1 and Samuel J. Cooper1

2025-04-24 0 0 8.35MB 10 页 10玖币

侵权投诉

MICROLIB: A LIBRARY OF 3D MICROSTRUCTURES

GENERATED FROM 2D MICROGRAPHS USING SliceGAN

Steve Kench 1, Isaac Squires 1, Amir Dahari 1, and Samuel J. Cooper 1

1Dyson School of Design Engineering, Imperial College London, London SW7 2DB

ABSTRACT

3D microstructural datasets are commonly used to deﬁne the geometrical domains used in ﬁnite element

modelling. This has proven a useful tool for understanding how complex material systems behave under

applied stresses, temperatures and chemical conditions. However, 3D imaging of materials is challenging

for a number of reasons, including limited ﬁeld of view, low resolution and difﬁcult sample preparation.

Recently, a machine learning method, SliceGAN, was developed to statistically generate 3D microstructural

datasets of arbitrary size using a single 2D input slice as training data. In this paper, we present the results

from applying SliceGAN to 87 different microstructures, ranging from biological materials to high-strength

steels. To demonstrate the accuracy of the synthetic volumes created by SliceGAN, we compare three

microstructural properties between the 2D training data and 3D generations, which show good agreement.

This new microstructure library both provides valuable 3D microstructures that can be used in models, and

also demonstrates the broad applicability of the SliceGAN algorithm.

Background

Understanding the inﬂuence of a material’s microstructure on its performance has led to signiﬁcant advancements

in the ﬁeld of material science

1–3

. Computational methods have played an important role in this success. For

example, ﬁnite element analysis can capture complex stress ﬁelds during mechanical deformation of structural

materials

4,5

, and electro-chemical modelling can help to explain rate limiting factors during battery discharge

6,7

These simulations allow high-throughput exploration of a systems performance under a range of conditions

8,9

. In

many ﬁelds, this has enabled massive acceleration of the materials optimisation process compared to experiments

alone, and with signiﬁcantly reduced cost. Importantly, 3D datasets are crucial for many applications where

2D datasets cannot be used to determine key material properties. For example, mechanical deformation, crack

propagation and tortuosity are three material characteristics that behave fundamentally differently in 3D compared

to 2D.

The ﬁdelity of the 3D microstructural datasets commonly required for physical modelling will inﬂuence the

simulations reliability. Unfortunately, to the authors knowledge, there are no 3D material databases, with most

data instead scattered across the literature. This is likely due to the high cost and technical experience required

for 3D imaging techniques, which inhibits free sharing of data. Furthermore, where there is data available, it

is commonly of limited resolution and ﬁeld of view due to the intrinsic 3D imaging constraints of techniques

such as focussed ion beam scanning electron microscopy and x-ray tomography

. In comparison, diverse, high

resolution 2D micrographs are abundantly available online due to the prevalence of 2D imaging techniques such

as light microscopy and scanning electron microscopy. DoITPoMS is one excellent micrograph repository with a

broad range of alloys, ceramics, bio-materials and more

. UHCSDB is a similar repository, focused solely on

high carbon steels

. ASM International has a collection of 4100 micrographs, though access costs a

250 yearly

subscription13.

arXiv:2210.06541v1 [cs.LG] 12 Oct 2022

KENCH et al. MICROLIB PREPRINT

In this paper, we aim to address the disparity between the availability of 2D micrographs compared to 3D. A

number of previous approaches have been developed to address this problem through dimensionality expansion,

which commonly entails statistical generation of 3D micrographs using statistics from a 2D training image.

These are typically physic based and require the extraction of particular metrics from the training data for

comparison. For example, sphere packing models using 2D particle size distributions

, poly-crystalline grain

growth algorithms15, and data fusion approaches16.

In this work, we use SliceGAN, a recently developed convolutional machine learning algorithm for dimensionality

expansion

. A typical GAN uses two convolutional networks (generator and discriminator) to learn to mimic

dataset distributions. The generator synthesises fake examples, and the discriminator identiﬁes differences

between these fake samples and the true training data distribution. Through iterative learning, the discriminator

informs the generator how to make increasingly realistic samples that match the real training data. Importantly,

in a typical set-up, the dimensionality of the generated images and the training data match. To facilitate different

dimensionalities, SliceGAN uses a simple modiﬁcation; a 3D generator network produces a sample volume,

then a 2D discriminator checks the ﬁdelity of one slice at a time, where the 2D dimensionality of the slice

now matches the 2D dimensionality of the training images. The algorithm is described in full in the original

manuscript

.SliceGAN is particularly well suited to the task at hand due to a number of key features. First,

broad applicability means that the same algorithm and hyper-parameters can be used for a very diverse set of

microstructures, as demonstrated in this dataset. Second, high speed training (typically 3 hrs on an RTX6000

GPU) and generation (

seconds for a 500

voxel volume) enables the synthesis of hundreds of large samples

for statistical experiments, as well as the generation of volumes far larger than it is currently possible to obtain

directly through imaging (

>20003

voxel). Third, complete automation of the 2D to 3D algorithm is possible

with no user deﬁned inputs, such as statistical features, being required. This combination of strengths makes

SliceGAN an excellent candidate for building the ﬁrst large scale 3D microstructural database from existing

open-source 2D data.

The beneﬁts of this database are twofold. First, we provide a diverse 3D microstructural dataset which can be

used by the material science community for modelling purposes. Crucially, users are not limited to the single

example cube we provide, as each data entry also has an associated trained generator neural network (45 Mb

in size) available to download. This can be used to synthesise arbitrary size datasets by cloning the SliceGAN

repo and running the relevant scripts (see methods). The second important function of this database is as a

demonstration to the material science community of the strengths of SliceGAN. The entries we provide are

diverse in their nature, and contained in an easily searchable website. Interested researchers can thus use this

website to check whether SliceGAN works on materials in their research ﬁeld, and see examples of generated

outputs. This encourages the submission of more entries to the database, and the further use of SliceGAN in the

ﬁeld of computational materials. The key data processing steps and datasets are presented in Figure 1.

Methods

As shown in Figure 1, the database construction required several distinct steps. First, a subset of micrographs

were selected using a set of exclusion criteria. A number of simple pre-processing operations were then applied

to ensure suitability for the SliceGAN workﬂow. An automated in-painting method was used to remove scale

bars from the micrographs; compared to a cropping approach, this saves crucial data in an already extremely

data-scarce setting. Finally, the resulting micrographs are used to train SliceGAN generators, each of which was

used to generate an example

3203

cubic volume. Each of these steps can be reproduced by cloning the MicroLib

repository and running main.py in the relevant modes, as described in the repository README.

Exclusion criteria

DoITPoMS includes 818 diverse micrographs which can easily be downloaded directly from their website.

However, not all are suitable for SliceGAN, which has a number of limitations. As such, the following exclusion

criteria are applied to leave 87 feasible microstructures:

Microstructure isotropy – SliceGAN can be used for some anisotropic microstructures, but this mode

requires multiple perpendicular micrographs which are not available from DoITPoMS.

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：

MICROLIB:ALIBRARYOF3DMICROSTRUCTURESGENERATEDFROM2DMICROGRAPHSUSINGSliceGANSteveKench1,IsaacSquires1,AmirDahari1,andSamuelJ.Cooper11DysonSchoolofDesignEngineering,ImperialCollegeLondon,LondonSW72DBABSTRACT3Dmicrostructuraldatasetsarecommonlyusedtodenethegeometricaldomainsusedinniteelementmodelling...

收起<<

MICRO LIB A LIBRARY OF 3D MICROSTRUCTURES GENERATED FROM 2D MICROGRAPHS USING SliceGAN Steve Kench1 Isaac Squires1 Amir Dahari1 and Samuel J. Cooper1.pdf

共10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！

MICRO LIB A LIBRARY OF 3D MICROSTRUCTURES GENERATED FROM 2D MICROGRAPHS USING SliceGAN Steve Kench1 Isaac Squires1 Amir Dahari1 and Samuel J. Cooper1

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: