面向高性能数值模拟的大数据分析共性技术研究

High performance numerical simulation is an important research tool for scientists to simulate complex problems at great fidelity and make science discoveries. With growing capability of supercomputers, scientists are able to simulate sophisticate physical phenomena with ever-increasing scale and accuracy. At the same time, thest numeric simulations lead to the generation of extremely lare amount of multidimension time-varying data. Confronting with such large scale scientific data, how to quickly carry out the data analysis, transfer data into knowledge, and then uncover inherent laws of physics is a key problem. Traditional data analysis methods have disadvantages such as low efficiency, hardware resources comsuming, and lacking of quantitative analysis. So, they can't meet the needs of practical applications. As a result, useful information is often overlooked. The project will deal with applications such as laser fusion,plasma physics, materials science, combining with scientific visualization, research large scale scientific data analysis common algorithms, and finally performe typical demonstrations. The main research content includes: (1)feature recognition technology, quickly locating physics moments and sub-regions which contain rich knowledge. (2)algorithms that improving visual expression. (3)quantitative analysis algorithms that support physical knowledge finding. (4)develop hiberarchy, module, scalble data analysis framework. (5)three or more typical application demonstrate the effectiveness of the proposed methods. The gains of the project can assist scientists to carry out data analysis, upgrade analysis efficiency, low difficulty, and remarkably improve the way in which scientists analyze and understand large scale scientific data.

高性能数值模拟是科学研究和大科学工程的重要研究手段，但同时也会产生TB以上量级的结果数据。面对如此大规模的科学数据，如何快速地开展数据分析，将数据转化为知识，进而揭示新的内在物理规律，成为科学研究的重要一环。传统的数据分析手段面临效率低、硬件资源要求高、缺少定量分析功能等挑战，不能充分满足实际应用需求。本项目面向激光聚变、等离子体物理、材料科学等应用领域，研究大数据分析过程中的共性算法，并进行典型示范。主要研究内容包括：（1）特征识别技术，快速定位蕴含丰富知识的物理时刻和局部计算子区域的算法；（2）可视分析技术，自动根据数据特征选择绘制参数，增强数据表现力的算法；（3）支撑物理知识发现的定量分析算法；（4）研制层次化、模块化、可扩展的数据分析框架；（5）三个以上应用领域的典型示范。项目成果可以辅助领域专家开展数据分析，提高分析效率，降低分析难度，对促进重大应用领域的科学发现具有重要意义。

本项目针对大规模数值模拟数据的特点和传统数据分析手段的缺点，从不同科学领域中抽取数据分析的共性支撑技术，形成可重用的算法模块，在不同数据分析场景中共享，从而提高数据分析的速度和效率，降低分析的难度。我们对标美国LLNL实验室相关研究，紧密联系数值模拟领域应用需求，与领域专家深入沟通交流，开展了卓有成效的合作研究工作。项目达成如下目标：（1）完成大规模科学数据的重要物理特征识别算法研究；（2）完成基于条件互信息的耦合变量因果关系分析方法研究；（3）完成可视化参数自动选择技术研究；（4）完成基于压缩感知的稀疏信号重构研究；（5）完成大规模分子动力学模拟数据中微结构识别算法研究；（6）完成大规模数值模拟数据分析软件初步研制。研究成果在物理分析、可视化和数据集之间建立一条通道，辅助科学家理解和分析物理规律，已经成为科学研究的关键支撑技术。相关成果已经应用于激光等离子体相互作用模拟、分子动力学模拟等数据的分析中。