3D Grain Shape Generation in Polycrystals Using Generative Adversarial Networks
3D Grain Shape Generation in Polycrystals Using Generative Adversarial Networks
复制标题
DOI：
10.1007/s40192-021-00244-1
复制
发表时间：
2022-01-21
影响因子：
3.3
通讯作者：
Manjunath, B. S.
中科院分区：
文献类型：
--
作者：
Jangid, Devendra K.;Brodnik, Neal R.;Manjunath, B. S.
本文提出了一种生成对抗网络（GAN），能够产生真实的微观结构形态特征，并在结晶钛颗粒形状数据集上展示了其能力。除此之外，我们提出了一种训练深度学习网络的方法，以基于与已建立的学习空间（如功能对象形状）的现有概念关系来理解特定于材料的描述符特征，如颗粒形状。具有Wasserstein损失的基于样式的GAN（称为M-GAN）首先被训练来识别ShapeNet数据集中功能对象的形态特征分布，然后应用于Ti-6Al-4V三维晶体数据集中的晶粒形态。对物体特征识别的评估显示出与最先进的基于体素的网络方法相当或更好的性能。当应用于实验数据时，M-GAN产生了与Ti-6Al-4V相媲美的真实晶粒形态。矩不变分布的定量比较表明，生成的颗粒在形状和结构上与地面真实相似，但从物体识别中学习的尺度不变性导致难以区分小颗粒的物理特征和空间分辨率伪像。讨论了M-GAN学习能力的物理含义，以及该方法在与晶粒形貌相关的其他材料特性中的可扩展性。
This paper presents a generative adversarial network (GAN) capable of producing realistic microstructure morphology features and demonstrates its capabilities on a dataset of crystalline titanium grain shapes. Alongside this, we present an approach to train deep learning networks to understand material-specific descriptor features, such as grain shapes, based on existing conceptual relationships with established learning spaces, such as functional object shapes. A style-based GAN with Wasserstein loss, called M-GAN, was first trained to recognize distributions of morphology features from function objects in the ShapeNet dataset and was then applied to grain morphologies from a 3D crystallographic dataset of Ti-6Al-4V. Evaluation of feature recognition on objects showed comparable or better performance than state-of-the-art voxel-based network approaches. When applied to experimental data, M-GAN generated realistic grain morphologies comparable to those seen in Ti-6Al-4V. A quantitative comparison of moment invariant distributions showed that the generated grains were similar in shape and structure to the ground truth, but scale invariance learned from object recognition led to difficulty in distinguishing between the physical features of small grains and spatial resolution artifacts. The physical implications of M-GAN's learning capabilities are discussed, as well as the extensibility of this approach to other material characteristics related to grain morphology.