本科技报告为国家重点研发计划“高通量自动流程材料集成计算算法与软件及其在先进存储材料中的应用”课题二“高通量材料数据管理、可靠性分析与智能学习”的验收报告。本课题主要研究目标包含：构建阻变/相变存储材料的第一原理计算高通量计算数据库系统；构建相变/阻变材料相图热力学数据库系统；分析与量化高通量计算材料数据的可靠性，并揭示其误差传递规律；实现海量异质材料数据的管理与挖掘开发面向新材料性能预测的智能学习与同化的集成化大数据处理模式。针对上述目标，课题组使用了数据库技术、不确定性量化等方法，在材料数据库、材料计算及共性算法、跨尺度计算和数据驱动的新材料设计这三个方面取得了重要成果。在数据库上，课题组自主构建了相变/阻变材料相图热力学数据库系统，并自主研发了材料数据管理系统Data Vault;开发了材料数据清洗、挖掘、分类技术，可实现与其它材料数据库平台的跨平台及连接查询的并行功能；利用项目组其它课题数据，共同构建了阻变/相变存储材料的第一原理计算高通量数据库系统。在共性算法上，课题组开发了可用于体模量、互扩散系数、二元相图等材料性质的高效算法，提供了动态系统状态预测修正的数据同化算法。在材料计算和新材料设计上，课题组计算了部分相变与阻变材料相关体系的相图；设计了可用于信息存储的二维本征铁磁体、Janus本征铁磁半导体和新型二维直接带隙半导体In2Ge2Te6。This report is the final report for program titled“High-throughput material data management,reliability analysis and intelligent learning”of the National Key Research and Development Program of China(high-throughput and automatic material integrated computation and software and applications in information-storage materials).The main objectives of this project are:construct the database for high-throughput ab initio data of phase-change RAM and resistance-change RAM materials;construct the database for phase diagrams of PC-RAM and R-RAM materials;analyze and quantify the reliability of high-throughput computational material data and explain its error propagation mechanism;achieve an integrated big-data process mode for large heterogeneous material data management and mining aiming for new material properties learning and data assimilation.To address those objectives,the program teams have employed multi-disciplinary methods and techniques including database and uncertainty quantification.In the duration of the program,a number of progresses have been made which can be categorized into three major fields:database,numerical frameworks for computational materials and similar problems,multi-scale computation and data-driven new material design.To be specific,for database,we have constructed the database for phase diagrams of PC-RAM and R-RAM materials;developed the material data management system Data Vault to provide internet services;developed material data cleaning,mining and category techniques and achieved parallel searching across various material database platforms;constructed the database for high-throughput ab initio data of PC-RAM and R-RAM materials using data from other participating teams of the same project.For numerical frameworks for computational materials and similar problems,we have developed efficient algorithms to estimate bulk modulus,inter-diffusion coefficients and phase diagrams,and developed data assimilation frameworks to correct prediction for dynamic systems。For multi-scale computation and data-driven new material design,we have computed the phase diagrams of phase-change RAM and resistance-change RAM related materials;design 2D ferromagnets,2D magnetic Janus semiconductor and 2D semiconductor In2Ge2Te6 using material data.

开发用户友好的高通量自动流程计算平台是材料基因工程思想实施和推广的重要基础，也是本课题的核心研究内容。本课题自主研发了高通量自动流程材料集成计算平台ALKEMIE,并部署于国家超算(天津、郑州)中心等5家超算中心;ALKEMIE可实现10000量级高通量自动流程计算，可实现建模、计算和分析的高通量自动流程。ALKEMIE具有跨尺度计算和机器学习功能，可实现第一原理高通量计算、机器学习和分子动力学模拟的无缝集成；实现了VASP与opencalphad及GIBBS2软件的耦合、VASP与动态蒙特卡洛的耦合；集成了相场计算软件Open Phase等。ALKEMIE具有友好的用户界面和良好的可扩展性、可移植性。基于ALKEMIE平台高通量计算筛选的新型相变存储材料已获得实验验证，包括Y-Sb2Te3、C-GeSb等相变存储材料。同时课题自主开发了材料性质计算专用算法和软件，包括电子结构计算算法shLDA-1/2、基于进化算法和团簇展开集成的缺陷结构预测算法pyGACE和金属性低维材料热导率的精准计算方法。在这些开发的计算平台、算法和软件的基础上，进行了多种材料的研究和设计，包括相变型信息存储材料、阻变型信息存储材料、其它新型信息存储材料等。同时还将这些方法拓展应用到了热电材料、新型能源材料等领域，展现了所研发方法和计算平台的普适性。The development of a user-friendly high-throughput automatic calculation platform is an important foundation for the material genome engineering,which is also the main content of this report.Here we developed the high-throughput automatic material calculation platform named ALKEMIE,and deployed it in 5 supercomputer centers including the National Supercomputer Center in Tianjin;ALKEMIE can perform 10000-level high-throughput automatic calculations.ALKEMIE has multi-scale calculation and machine learning functions,it has integrated DFT with MD,VASP with opencalphad,DFT with KMC.ALKEMIE has a friendly user interface and good scalability.New phase-change information storage materials screened based on the high-throughput calculation using ALKEMIE have been experimentally verified,including Y-Sb2Te3 and C-GeSb.At the same time,we developed algorithm and software for material properties calculation,including electronic structure calculation algorithm shLDA-1/2,defect structure prediction algorithm pyGACE based on evolutionary algorithm and cluster expansion method,and accurate calculation method for thermal conductivity of metallic low-dimensional materials.On the basis of these developed computing platform,algorithm and software,research and design of a variety of materials have been carried out,including phase change information storage materials,resistance switching information storage materials,and other new information storage materials.In addition to the materials for information storage,these methods are also applied to thermoelectric materials,energy materials and other fields.

本课题“先进存储材料与器件的全流程设计及制备”由中科院上海微系统与信息技术研究所、中科院微电子研究所和北京航空航天大学共同承担，负责相变和阻变存储材料与器件的设计、制备、性能验证与机理研究等方面的工作。主要研究内容：基于相变和阻变材料制备、表征以及存储器加工、测试平台，从速度、功耗、循环次数、数据保持力等关键性能方面验证由高通量计算获得的新型相变和阻变材料应用于相变和阻变存储器的可行性；结合高分辨微结构表征手段、理论计算与模拟、电性能测试结果研究相变和阻变材料及其器件的存储机理。达到的研究目标：通过3-4年的研究，研制出2-3种优化组分的存储材料及其原型器件，为加快我国相变和阻变存储技术的产业化发展提供支撑和指导。实现的主要技术指标：实现器件存储窗口一个数量级以上，操作速度达10 ns,擦写次数达1E6次。The project"Full-process design and preparation of advanced materials and devices for data storage"is jointly undertaken by Shanghai Institute of Microsystem and Information Technology,Chinese Academy of Sciences,Institute of Microelectronics,Chinese Academy of Sciences,and Beihang University.It is responsible for phase change and resistance change storage materials and devices.Design,preparation,performance verification and mechanism research work.Main research content:Based on the preparation and characterization of phase change and resistance change materials,as well as memory processing and test platforms,verify the new phase change and resistance obtained by high-throughput calculations in terms of speed,power consumption,cycle times,data retention and other key performance The feasibility of the application of variable materials in phase change and resistive memory;combining high-resolution microstructure characterization methods,theoretical calculations and simulations,and electrical performance test results to study the storage mechanism of phase change and resistive materials and their devices.Research goal reached:Through 3-4 years of research,develop 2-3 kinds of optimized composition storage materials and their prototype devices to provide support and guidance for accelerating the industrialization of phase change and resistive storage technology in my country.Main technical indicators achieved:Achieve a storage window of more than one order of magnitude,an operating speed of 10 ns,and a cyclic erasing and writing times of 1E6.

The catering big data of large cities are typically featured by spatial and temporal characteristics. It is a general requirement, applying to a range of data mining technologies, to discover the underlying features of restaurant data sets. Here we conduct a data-mining study of Dianping restaurant data in Shanghai of more than 30K records. The spatial location of the most popular cuisines is demonstrated to be of concentrated distribution. Considering the price and review scores of the restaurants, we show that both these two properties have strong spatial correlations characterized by the Moran's I index. To further analyze the temporal characteristics of the catering data, we make a more detailed spatial-temporal joint discussion and construct temporal networks for spatial neighboring restaurants for 14 years. Regression and statistical analysis suggest that the catering industry of Shanghai has become negative assortativity and has been significantly improved in connectivity and scales. © 2020 IEEE.

PurposeMetamodeling is an effective method to approximate the relations between input and output parameters when significant efforts of experiments and simulations are required to collect the data to build the relations. This paper aims to develop a new sequential sampling method for adaptive metamodeling by using the data with highly nonlinear relation between input and output parameters.Design/methodology/approachIn this method, the Latin hypercube sampling method is used to sample the initial data, and kriging method is used to construct the metamodel. In this work, input parameter values for collecting the next output data to update the currently achieved metamodel are determined based on qualities of data in both the input and output parameter spaces. Uniformity is used to evaluate data in the input parameter space. Leave-one-out errors and sensitivities are considered to evaluate data in the output parameter space.FindingsThis new method has been compared with the existing methods to demonstrate its effectiveness in approximation. This new method has also been compared with the existing methods in solving global optimization problems. An engineering case is used at last to verify the method further.Originality/valueThis paper provides an effective sequential sampling method for adaptive metamodeling to approximate highly nonlinear relations between input and output parameters.

This paper presents a novel method to generate three-axis CNC tool paths for machining freeform surfaces based on a preferred feed direction field. This research was initiated from a fluid dynamics behavior that the energy loss can be reduced when the streamlines of fluid and the small grooves on a surface are in the same directions. In this research, the fluid streamlines above the surface are defined by a collection of vectors. These vectors are regularized into a grid of vectors, and these regularized vectors are further projected onto the tangent planes of a grid of points on the surface to create the preferred feed direction field. Based on the parametric model of the surface, the vectors on the tangent planes of the surface are mapped into vectors in the parametric domain. A scalar function is constructed such that the isolines of this scalar function and the preferred feed direction vectors in the parametric domain are in the same directions. A group of isolines of the scalar function are identified and these isolines are mapped back onto the 3-D surface as the created tool paths considering the tolerance requirement. The developed method has been applied to generate the tool paths for machining surfaces of a compressor blade. © 2019, Korean Society for Precision Engineering.

Financial price series trend prediction is an essential problem which has been discussed extensively using tools and techniques of economic physics and machine learning. Time dependence and volatility issues in this problem have made Hidden Markov Model (HMM) a useful tool in predicting the states of stock market. In this paper, we present an approach to predict the stock market price trend based on high-order HMM. Different from the commonly used first-order HMM, short and long-term time dependence are both considered in the high order HMM. By introducing a dimension reduction method which could transform the high-dimensional state vector of high-order HMM into a single one, we present a dynamic high-order HMM trading strategy to predict and trade CSI 300 and S&P 500 stock index for the next day given historical data. In our approach, we make a statistic of the daily returns in the history to demonstrate the relationship between hidden states and the price change trend. Experiments on CSI 300 and S&P 500 index illustrate that high-order HMM has preferable ability to identify market price trend than first-order one. Thus, the high-order HMM has higher accuracy and lower risk than the first-order model in predicting the index price trend. © 2018 Elsevier B.V.

Understanding the correlations among stock returns is crucial for reducing the risk of investment in stock markets. As an important stylized correlation, lead–lag effect plays a major role in analyzing market volatility and deriving trading strategies. Here, we explore historical lead–lag relationships among stocks in the Chinese stock market. Strongly positive lagged correlations can be empirically observed. We demonstrate this lead–lag phenomenon is not constant but temporally emerges during certain periods. By introducing moving time window method, we transform the lead–lag dynamics into a series of asymmetric lagged correlation matrices. Dynamic lead–lag structures are uncovered in the form of temporal network structures. We find that the size of lead–lag group experienced a rapid drop during the year 2012, which signaled a re-balance of the stock market. On the daily timescale, we find the lead–lag structure exhibits several persistent patterns, which can be characterized by the Jaccard matrix. We show significant market events can be distinguished in the Jaccard matrix diagram. Taken together, we study an integration of all the temporal networks and identify several leading stock sectors, which are in accordance with the common Chinese economic fundamentals. © 2018 Elsevier B.V.

As the extensive use of first-principles Density Functional Theory (DFT) simulations, using DFT for high-throughput screening to predict the desirable doped structures that are physically stable with optimal properties becomes common. Usually, the challenge of running doping calculation is how to obtain inequivalent doped structures as input for DFT simulations to find the desirable doped structure(s). The current practice of substitutional doping is mainly based on experience to use one or more dopant atoms to replace target atoms to be substituted. Using this manual approach to produce all inequivalent doped structures based on expertise is tedious, and the results are usually incomplete. To address this need, we propose a "doping-filtering" collaboratively working approach and develop associated high-throughput computational algorithms to obtain inequivalent doped structures for substitutional doping-based high-throughput screening effectively. A computational time benchmark matrix table of using this approach to obtain inequivalent doped structures for different doping concentrations is also given. The algorithm is integrated into a high-throughput computational material infrastructure named MatCloud. It has been demonstrated in the study case of doping Ni, Co, Ti and Sc into Zr2Fe that the approach and algorithm are effective in reducing the computational time in obtaining inequivalent doped structures. © 2018 Elsevier B.V.