ITR: Simulating Extended Time and Length Scales using Parallel Kinetic Monte Carlo and Parallel Accelerated Dynamics

该奖项是提交给信息技术研究（ITR）倡议的提案的结果。这项研究将与洛斯阿拉莫斯国家实验室合作完成。理解凝聚相系统的一个长期障碍是，许多重要过程发生在一个时间尺度上，而传统的模拟方法不容易获得。例如，分子动力学通常限于纳秒，因为积分运动方程所需的时间步长很小。然而，被激活的相关过程，即不频繁的事件，通常在微秒甚至更长时间内发生。例子包括晶体或薄膜生长过程中表面形貌的演变，固体中点缺陷的扩散，以及塑性应变过程中晶界的迁移。近年来，人们提出了各种加速动力学技术，包括超动力学、平行复制动力学和温度加速动力学，以加快对分子动力学中罕见事件的模拟。特别是，温度加速动力学在延长模拟的时间尺度方面非常成功，因为它允许在时间尺度上对低温过程进行真实的模拟，长至几秒甚至几小时。然而，由于计算工作需要的尺度是原子数N的平方甚至立方，这种技术只能应用于极小的系统。因此，在延长的时间和延长的长度尺度上对材料进行真实的模拟是不可能的。这里的研究将使用并行计算来扩展加速动力学，以便同时模拟大系统尺寸和长时间尺度。这种模拟介观系统大小和长时间尺度的能力的发展应该是我们进行真实原子模拟能力的重大突破。与此同时，将开发稀疏计算算法，以减少加速动力学中计算工作量对系统大小的依赖所对应的指数。这种算法的发展是基于这样一种认识，即在搜索鞍点期间，从这些方法中涉及的力计算中消除非局部运动或非局部运动组应该显著减少对群集大小n所需的计算工作的依赖。并行复制动力学也将用于这两种技术中，以延长模拟的时间尺度。作为空间平行加速动力学的第一个具体应用，我们将重点关注使用温度加速动力学在低温下模拟金属对金属外延生长。这是一个非常有趣的问题，因为在低温下观察到各种无法解释的现象，包括纳米尺度的面化和应变诱导的丘正则化。这里开发的方法也应该适用于更广泛的系统。该合同是提交给信息技术研究（ITR）倡议的一项提案的结果。这项研究将与洛斯阿拉莫斯国家实验室合作完成。理解凝聚相系统的一个长期障碍是，许多重要过程发生在一个时间尺度上，而传统的模拟方法不容易获得。例如，分子动力学通常限于纳秒，因为积分运动方程所需的时间步长很小。然而，被激活的相关过程，即不频繁的事件，通常在微秒甚至更长时间内发生。例子包括晶体或薄膜生长过程中表面形貌的演变，固体中点缺陷的扩散，以及塑性应变过程中晶界的迁移。近年来，人们提出了各种加速动力学技术，包括超动力学、平行复制动力学和温度加速动力学，以加快对分子动力学中罕见事件的模拟。特别是，温度加速动力学在延长模拟的时间尺度方面非常成功，因为它允许在时间尺度上对低温过程进行真实的模拟，长至几秒甚至几小时。然而，由于计算工作需要的尺度是原子数N的平方甚至立方，这种技术只能应用于极小的系统。因此，在延长的时间和延长的长度尺度上对材料进行真实的模拟是不可能的。这里的研究将使用并行计算来扩展加速动力学，以便同时模拟大系统尺寸和长时间尺度。这种模拟介观系统大小和长时间尺度的能力的发展应该是我们进行真实原子模拟能力的重大突破。与此同时，将开发稀疏计算算法，以减少加速动力学中计算工作量对系统大小的依赖所对应的指数。这种算法的发展是基于这样一种认识，即在搜索鞍点期间，从这些方法中涉及的力计算中消除非局部运动或非局部运动组应该显著减少对群集大小n所需的计算工作的依赖。并行复制动力学也将用于这两种技术中，以延长模拟的时间尺度。作为空间平行加速动力学的第一个具体应用，我们将重点关注使用温度加速动力学在低温下模拟金属对金属外延生长。这是一个非常有趣的问题，因为在低温下观察到各种无法解释的现象，包括纳米尺度的面化和应变诱导的丘正则化。这里开发的方法也应该适用于更广泛的系统

This award is the result of a proposal submitted to the Information Technology Research (ITR) initiative. The research will be done in collaboration with Los Alamos National Laboratory. A long-standing obstacle to the understanding of condensed phase systems is that many important processes occur on a time-scale that is not easily accessible with conventional simulation methods. For example, molecular dynamics is generally limited to nanoseconds because of the small time-step required for the integration of the equations of motion. However, relevant processes that are activated, i.e., infrequent events, often take place on a time-scale of microseconds or even longer. Examples include the evolution of the surface morphology during crystal or film growth, the diffusion of point defects in solids, and the migration of grain boundaries during plastic strain.Recently, a variety of accelerated dynamics techniques, including hyperdynamics, parallel replica dynamics, and temperature-accelerated dynamics have been proposed in order to speed up the simulation of infrequent events in molecular dynamics. In particular, temperature-accelerated dynamics has been quite successful in extending the time-scales for simulations since it allows realistic simulations oflow-temperature processes over time-scales as long as seconds and even hours. However, due to the fact that the computational work required scales as the square or even as the cube of the number of atoms N, this technique can only be applied to extremely small systems. As a result, realistic simulations of materials over both extended time and extended length scales have not been possible.The research here will use parallel computations in order to extend accelerated dynamics so that both large system sizes and long time-scales can be simulated simultaneously. The development of such a capability to simulate both mesoscopic systems sizes and long time-scales should present a major breakthrough in our ability to carry out realistic atomic simulations.In parallel with this effort, sparse-computational algorithms will be developed in order to reduce the exponent corresponding to the dependence of the computational work on the system size in accelerated dynamics. The development of such algorithms is based on the realization that eliminating non-local moves or groups of non-local moves from the force calculations involved in these methods during the search for saddle-points should significantly reduce the dependence of the computational work required on the cluster size N. Parallel replica dynamics will also be used in both techniques to extend the time-scales of the simulations.As a first specific application of spatially parallel accelerated dynamics we will focus on simulations of metal-on-metal epitaxial growth at low temperature using temperature-accelerated dynamics. This is a problem of great interest due to the observation of a variety of unexplained phenomena including nanoscale facetting and strain-induced mound regularization at low tempertures. The methods developed here should be applicable to a much broader range of systems as well.%%% This award is the result of a proposal submitted to the Information Technology Research (ITR) initiative. The research will be done in collaboration with Los Alamos National Laboratory. A long-standing obstacle to the understanding of condensed phase systems is that many important processes occur on a time-scale that is not easily accessible with conventional simulation methods. For example, molecular dynamics is generally limited to nanoseconds because of the small time-step required for the integration of the equations of motion. However, relevant processes that are activated, i.e., infrequent events, often take place on a time-scale of microseconds or even longer. Examples include the evolution of the surface morphology during crystal or film growth, the diffusion of point defects in solids, and the migration of grain boundaries during plastic strain.Recently, a variety of accelerated dynamics techniques, including hyperdynamics, parallel replica dynamics, and temperature-accelerated dynamics have been proposed in order to speed up the simulation of infrequent events in molecular dynamics. In particular, temperature-accelerated dynamics has been quite successful in extending the time-scales for simulations since it allows realistic simulations oflow-temperature processes over time-scales as long as seconds and even hours. However, due to the fact that the computational work required scales as the square or even as the cube of the number of atoms N, this technique can only be applied to extremely small systems. As a result, realistic simulations of materials over both extended time and extended length scales have not been possible.The research here will use parallel computations in order to extend accelerated dynamics so that both large system sizes and long time-scales can be simulated simultaneously. The development of such a capability to simulate both mesoscopic systems sizes and long time-scales should present a major breakthrough in our ability to carry out realistic atomic simulations.In parallel with this effort, sparse-computational algorithms will be developed in order to reduce the exponent corresponding to the dependence of the computational work on the system size in accelerated dynamics. The development of such algorithms is based on the realization that eliminating non-local moves or groups of non-local moves from the force calculations involved in these methods during the search for saddle-points should significantly reduce the dependence of the computational work required on the cluster size N. Parallel replica dynamics will also be used in both techniques to extend the time-scales of the simulations.As a first specific application of spatially parallel accelerated dynamics we will focus on simulations of metal-on-metal epitaxial growth at low temperature using temperature-accelerated dynamics. This is a problem of great interest due to the observation of a variety of unexplained phenomena including nanoscale facetting and strain-induced mound regularization at low tempertures. The methods developed here should be applicable to a much broader range of systems as well.***