【作者】 王鹏；

【导师】 梅金硕；

【作者基本信息】 哈尔滨理工大学 ， 电子科学与技术， 2021， 硕士

【摘要】 目前,对于新材料的探索往往是在经验指导下通过不断尝试来寻求更优良的性能,存在一定的偶然性。在计算材料固有属性上第一性原理计算仍然是最为主流的计算方式。但是这一计算方法的实现需要进行复杂的准备工作和消耗大量的运算时间,甚至对于计算软件操作的专业性也有较高要求。近年来,机器学习算法因其优秀的处理大数据量计算效率和预测能力正在越来越多地被应用到探索新材料和计算材料属性领域,逐渐展现出其革命性的优势,博得了业内研究者的极大关注。因此,基于机器学习算法加速材料固有属性的第一性原理计算具有十分重要的研究意义。针对Zr-O二元体系不同构型下形成能的计算,本文首先基于团簇扩展方法进行机器学习特征选取,采用合适的团簇函数以区分不同的原子有序度,即不同的构型可由不同的团簇函数进行描述,并使用团簇相关函数作为指纹区分对称不同的团簇轨道。利用这一方法获得了一个具有多体相互作用的合成团簇扩展哈密顿量,并根据第一性原理计算出晶体在不同构型情况下的形成能。此外,本文在团簇扩展方法的基础上,忽略团簇相关函数与构型能之间的线性约束,通过机器学习模型建立一种团簇相关函数与构型能之间的非线性映射关系。将团簇相关函数用作神经网络、高斯过程回归等回归算法模型的输入特征,通过第一性原理给出的数据集训练模型进行构型能预测,加速构型能计算。本文基于四种机器学习算法结合团簇相关函数和间隙位状态集合这两类特征进行模型训练和测试。其中,对团簇相关函数特征使用遗传算法结合团簇扩展公式、递归特征消除结合支持向量回归、随机森林等四种优化方案进行特征优化。进而对所有加速构型能计算方案的预测结果进行对比分析。最终分析结果表明,采用团簇扩展公式作为遗传算法评估模型优化特征,并结合高斯过程回归模型进行预测为最优加速构型能计算方案。更多还原

【Abstract】 At present,the exploration of new materials is often based on continuous attempts to seek better performance under the guidance of experience,and there is a few of contingency.First-principles calculation is still the most mainstream method for calculating the inherent properties of materials.However,the realization of this calculation method requires complicated preparation work and consumes a lot of calculation time,and even has higher requirements for the professionalism of calculation software operations.In recent years,machine learning algorithms are increasingly being applied to the field of exploring new materials and calculating material properties due to their excellent computing efficiency and predictive capabilities for processing large amounts of data,gradually showing their revolutionary advantages and winning great attention of the industry insiders.Therefore,first-principles calculations based on machine learning algorithms to accelerate the inherent properties of materials have very important research significance.Aiming at the calculation of the formation energies of the Zr-O binary system under different configurations,this paper first selects machine learning features based on the cluster expansion method,and uses appropriate cluster functions to distinguish different atomic order degrees,that is,different configurations can be described by different cluster functions,and cluster correlation functions are used as fingerprints to distinguish clusters with different symmetry orbitals.Using this method,a synthetic cluster with multi-body interactions with extended Hamiltonian was obtained,and the formation energy of the crystal in different configurations was calculated according to the first principles.In addition,based on the cluster expansion method,this paper ignores the linear constraint between the cluster correlation function and the configuration energy,and establishes a nonlinear mapping relationship between the cluster correlation function and the configuration energy by a machine learning model.The cluster correlation function is used as the input feature of regression algorithm models such as neural network and Gaussian process regression,and the configurational energy prediction is carried out through the data set training model given by the first principles,and the configurational energy calculation is accelerated.This paper is based on four machine learning algorithms combining cluster correlation functions and gap state sets for model training and testing.Among them,four optimization schemes including genetic algorithm combined with cluster expansion formula,recursive feature elimination combined with support vector regression and random forest are used to optimize the features of cluster correlation functions.Furthermore,the prediction results of all accelerated the configurational energy calculation schemes are compared and analyzed.The final analysis results show that the cluster expansion formula is used as the genetic algorithm to evaluate the optimization characteristics of the model,and combined with the Gaussian process regression model to predict the optimal acceleration configuration energy calculation scheme.更多还原