海归学者发起的公益学术平台
分享信息，整合资源
交流学术，偶尔风月
近来，神经网络和高斯过程回归等机器学习工具越来越多地应用于原子作用势的相关研究。
来自加州大学圣巴巴拉分校的AntonVan der Ven领导的团队，利用机器学习方法亦可预测晶体形成能。他们开发了一种先进的神经网络方法，借助适度数量的关联函数作为描述符，构建了精确的格点哈密顿模型，来描述多组分固体中依赖格点位置占据几率的性质。利用位点中心关联函数作为描述符，该方法精确地得到面心立方晶体的综合多体二元哈密顿函数的形成能，以及锂插层TiS2的形成能。结果表明，复杂的多体相互作用可由非线性模型来近似描述，该描述借助较小的集团即可获得。该方法可以进一步拓展用于描述多组分晶体中给定构型自由度下任意的标量性质（包括形成能和体积）。
该文近期发表于npj Computational Materials 4: 56 (2018)，英文标题与摘要如下，点击左下角“阅读原文”可以自由获取论文PDF。
Machine-learning the configurational energy of multicomponent crystalline solids
Anirudh Raju Natarajan & Anton Van der Ven
Machinelearning tools such as neural networks and Gaussian process regression areincreasingly being implemented in the development of atomistic potentials. Here,we develop a formalism to leverage such non-linear interpolation tools indescribing properties dependent on occupation degrees of freedom inmulticomponent solids. Symmetry-adapted cluster functions are used todifferentiate distinct local orderings. These local features are used as inputto neural networks that reproduce local properties such as the site energy. Weapply the technique to reproduce a synthetic cluster expansion Hamiltonian withmulti-body interactions, as well as the formation energies calculated fromfirst-principles for the intercalation of lithium into TiS2. Theformalism and results presented here show that complex multi-body interactionsmay be approximated by non-linear models involving smaller clusters.