MSPA-MCS: Data-Driven Parallelization of Time in Molecular Dynamics Simulations

MSPA-MCS：分子动力学模拟中数据驱动的时间并行化

• 批准号: 0626180
• 负责人: Ashok Srinivasan
• 金额: --
• 依托单位: Florida State University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2006
• 资助国家: 美国
• 起止时间: 2006-09-01 至 2010-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0626180&HistoricalAwards=false
• 关键词: MSPA; MCS; Data; Driven; Parallelization

Conventional parallelization strategies do not scalewell when the computational effort arises from theneed to simulate to long time spans, rather than fromlarge state space. Molecular Dynamics simulationsconstitute an important class of applications wherethis proves to be a bottleneck. The investigatorsdevelop a new approach to parallelization of MolecularDynamics, which is based on the observation thatsimulations typically occur in a context rich in datafrom other related simulations. They use such data toparallelize the time domain, which yields a morescalable algorithm. This approach is based on theobservation that long time-spans are often encounteredin simulations with multiple time scales. The finescales are responsible for the large computationaleffort. However, the important contribution of thefine-scales is often to the effect they have on thecoarse scale. The investigators use reduced ordermodeling to identify important coarse scale effects.They use clustering and machine learning todynamically determine the relationship between thesimulation being performed and prior data. They useODE and controls theory for stability analysis,uncertainty estimation, and system identification.The investigators validate their techniques using avariety of realistic applications in nano and bio-nanomaterials. The importance of the applications chosenarises from the fact that materials have historicallyplayed a pivotal role in human progress. An indicationof their importance lies in the fact that eras ofhuman progress, such as the iron age and the bronzeage, are named after the materials that contributed tosuch progress. Nano and bio-nano materials, designedbased on fundamental understanding at the atomicscale, promise yet another revolution, leading toproducts such as fuel efficient cars, disasterresistant structures, and new ways of treatingdiseases. The investigators' techniques remove animportant impediment to such developments.Furthermore, the students involved obtain training ininterdisciplinary research. Inclusion of acollaborator with a joint appointment at an HBCUensures that under-represented students too benefitfrom this work.

当计算工作是由于需要模拟较长的时间跨度而不是来自较大的状态空间时，传统的并行化策略无法很好地扩展。分子动力学模拟构成了一类重要的应用，这被证明是一个瓶颈。研究人员开发了一种分子动力学并行化的新方法，该方法基于这样的观察：模拟通常发生在富含其他相关模拟数据的环境中。他们使用此类数据来并行化时域，从而产生更具可扩展性的算法。这种方法基于这样的观察：在多时间尺度的模拟中经常遇到长时间跨度。精细尺度导致大量计算工作。然而，细尺度的重要贡献往往在于它们对粗尺度的影响。研究人员使用降阶建模来识别重要的粗尺度效应。他们使用聚类和机器学习来动态确定正在执行的模拟与先前数据之间的关系。他们使用常微分方程和控制理论进行稳定性分析、不确定性估计和系统识别。研究人员使用纳米和生物纳米材料中的各种实际应用来验证他们的技术。所选应用的重要性源于这样一个事实：材料历来在人类进步中发挥着关键作用。它们的重要性在于，人类进步的时代，例如铁器时代和青铜器时代，是以促进这种进步的材料命名的。基于原子尺度的基本理解而设计的纳米和生物纳米材料有望带来另一场革命，带来诸如节能汽车、抗灾结构和治疗疾病的新方法等产品。研究人员的技术消除了此类发展的重要障碍。此外，参与的学生获得了跨学科研究的培训。在 HBCU 联合任命一位合作者可以确保代表性不足的学生也能从这项工作中受益。