ITR/AP(DMR): Billion-Atom Multiscale Simulations of Nanosystems on a Grid

该奖项是提交给信息技术研究计划的提案的结果。这项研究的目标是开发一种可扩展的软件基础设施，用于在地理分布的网格上进行大型多尺度模拟，大规模并行超级计算机，以及未来的千万亿次计算机。多尺度模拟方法将在单一网格软件中结合有限元（FE）计算、粗粒度分子动力学（CGMD）、分子动力学（MD）模拟和基于密度泛函理论（DFT）的量子力学（QM）计算。基于有限元方法的连续介质力学计算将与CGMD方法导出的本构关系结合MD模拟进行，而MD模拟将嵌入由DFT描述的QM算法。将开发以下内容：(1)基于网格的FE/CGMD/MD/QM算法，该算法基于在并行/分布式计算机上实现分层分解的时空多分辨率算法，以实现可扩展性；(2)结合动力学蒙特卡罗（KMC）和并行复制方法的时空分割多尺度模拟，将不同的长度和时间尺度耦合起来；(3)网格计算工具，包括基于小波计算空间分解的自适应负载均衡和基于空间填充曲线的自适应数据压缩，以减少通信和存储；(4)利用基于八叉树的可视性剔除技术实现大型仿真数据的沉浸式交互式可视化，并利用机器学习预测预取技术对可视化数据进行并行/分布式预处理。栅格化软件将用于研究对未来信息处理具有重要意义的纳米系统。将进行涉及1,000 - 10,000个QM原子和1亿- 10亿个MD原子的多尺度模拟，以研究原子诱导现象，重点是化学过程发挥重要作用的环境影响。多尺度算法将原子过程与实验观察到的量联系起来，通过连续介质力学覆盖一个数量级更大的长度尺度（10微米），并通过KMC和复制方法扩展时间尺度。模拟将集中在用于亚0.1微米微电子应用的Si/Si3N4和GaAs/Si3N4纳米像素中的应力域及其声子成像及其氧化效应，以及衬底编码的晶格不匹配半导体量子点（GaAs/InAs）的自组织生长。该项目将涉及与政府实验室（Argonne, NASA Ames, Sandia, Naval Oceanographic Office），工业界（Intel, Motorola）和大学的科学家以及欧洲，日本和南美的国际合作的密切合作

This award is the result of a proposal submitted to the Information Technology Research initiative. The goal of the research is to develop a scalable software infrastructure for large multiscale simulations on a Grid of geographically distributed, massively parallel supercomputers, as well as on future Petaflop computers.The multiscale simulation approach will combine, in a single Grid software, finite element (FE) calculation, the coarse-grained molecular dynamics (CGMD), molecular dynamics (MD) simulation, and quantum mechanical (QM) calculation based on the density functional theory (DFT). Continuum mechanics calculation based on the FE method will be performed with constitutive relations derived from the CGMD method in conjunction with MD simulations, which in turn will embed QM algorithm described by the DFT. The following will be developed: (1) Grid-based FE/CGMD/MD/QM algorithms based on space-time multiresolution algorithms implemented with hierarchical decomposition on parallel/distributed computers for scalability and constrained-dynamics-based hybridization for seamless coupling of the hybrid simulation componets; (2) Space-time partitioned multiscale simulation combined with kinetic Monte Carlo (KMC) and parallel replica methods to couple disparate length and time scales; (3) Grid-computation tools including adaptive load balancing using wavelet-based computational-space decomposition and space-filling-curve-based adaptive data compression to reduce communication and storage; (4) Immersive and interactive visualization of the large simulation data using octree-based visibility culling and parallel/distributed preprocessing of the visualization data with machine-learning predictive prefetch.The Gridified software will be used to study nanosystems of great importance to future information processing. Multiscale simulations involving 1,000 - 10,000 QM atoms and 100 million - 1 billion MD atoms will be performed to study atomistically-induced phenomena, with emphasis on environmental effects where chemical processes play an important role. The multiscale algorithm will relate the atomistic processes to experimentally observable quantities, by covering an order-of-magnitude larger length scale (10 micron) through continuum mechanics and extending time scales through the KMC and replica methods. The simulations will focus on stress domains and their phonon imaging in Si/Si3N4 and GaAs/Si3N4 nanopixels for sub-0.1 micron microelectronics applications and oxidation effects on them, and on substrate-encoded self-organized growth of lattice-mismatched semiconductor quantum dots (GaAs/InAs). The project will involve close collaborations with scientists at government laboratories (Argonne, NASA Ames, Sandia, Naval Oceanographic Office), industry (Intel, Motorola) and universities, as well as international collaborations in Europe, Japan and South America.%%%