arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2101.01773cond-mat.mtrl-scicond-mat.str-el
收藏

学习晶体结构基因组用于属性分类

Learning the Crystal Structure Genome for Property Classification

Yiqun Wang, Xiao-Jie Zhang, Fei Xia, Elsa A. Olivetti, Ram Seshadri, James M. Rondinelli

摘要:

材料属性预测从机器学习算法的进步中得到了改进,通过结构-属性关系的数据驱动模型提供材料发现和新见解。几乎所有可用的模型都依赖于材料成分的特征化,然而,在这种模型中是否专门使用结构知识进行可比预测的能力仍然是未知的。在这里,我们采用深度神经网络模型来解码晶体材料中的结构-性能关系,而无需明确考虑化学成分。重点是晶体系统的分类,机械弹性,电子带隙,和相稳定性。我们的模型采用了三维(3D)动量空间表示的结构,从弹性X射线散射理论,表现出旋转和排列不变性。我们进行了新的消融研究,以帮助解释模型的性能,通过扰动物理上有意义的输入特征(即,衍射图案),而不是像传统烧蚀方法那样调整学习模型的结构。我们发现,三维衍射图案的空间对称性,这反映了晶体的对称操作,是更重要的比衍射强度内的模型,使一个成功的分类。我们的工作展示了使用统计学习模型来帮助理解材料物理的潜力,而不是像大多数材料信息学研究那样执行预测和生成任务。我们还认为,以化学不可知的方式学习晶体结构基因组表明,一些晶体结构固有地具有最佳材料性能的高倾向性,这使得结构和组成的解耦能够用于未来的多功能协同设计。

英文摘要:

Materials property predictions have improved from advances in machine learning algorithms, delivering materials discoveries and novel insights through data-driven models of structure-property relationships. Nearly all available models rely on featurization of materials composition, however, whether the exclusive use of structural knowledge in such models has the capacity to make comparable predictions remains unknown. Here we employ a deep neural network model to decode structure-property relationships in crystalline materials without explicitly considering chemical compositions. The focus is on classification of crystal systems, mechanical elasticity, electronic band gap, and phase stability. Our model utilizes a three-dimensional (3D) momentum space representation of structure from elastic x-ray scattering theory that exhibits rotation and permutation invariance. We perform novel ablation studies to help interpret the model performance by perturbing the physically meaningful input features (i.e., the diffraction patterns) instead of tuning the architecture of the learning model as in conventional ablation methods. We find that the spatial symmetry of the 3D diffraction patterns, which reflects crystalline symmetry operations, is more important than the diffraction intensities contained within for the model to make a successful classification. Our work showcases the potential of using statistical learning models to help understand materials physics, rather than performing predictive and generative tasks as in most materials informatics research. We also argue that learning the crystal structure genome in a chemistry-agnostic manner demonstrates that some crystal structures inherently host high propensities for optimal materials properties, which enables the decoupling of structure and composition for future codesign of multifunctionality.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑