Accelerated Dynamics of Surface Chemical Reactions

1600218 PI：Ramprasad研究所：康涅狄格大学标题：表面化学反应的加速动力学表面化学反应在自然现象中普遍存在，在多相催化、晶体生长和电化学等几个科学学科中发挥着关键作用。然而，要在分子水平上了解表面化学反应的机理和动力学细节，仍然是一项艰巨的任务。目前的研究方法，无论是经验性的还是计算性的，都只为我们提供了对这个复杂世界的有限视角。达到必要的理解水平的一种方法是使用分子动力学(MD)模拟。这些模拟可以直接在分子水平上监测表面反应的进程，作为各种相关条件的函数。然而，在当今的MD能力中仍然存在着显著的差距：它们要么是快速的(但不是通用的或准确的)，例如基于经验力场的能力，要么是通用和准确的(但不是有效的)，例如基于量子力学(或从头算)方法的能力。这项拟议的工作将利用一种自适应机器学习方案，既可以加速从头算MD模拟，又可以动态创建准确的力场(不需要额外成本)。以前用量子力学模拟无法达到的时间尺度可以使用这种新的范例(可以想象，几毫秒到几秒)来访问，同时仍然保持量子力学的保真度。从头算MD模拟速度慢的主要原因在很大程度上是因为当今范例中渗透的巨大冗余。无论新配置是否与先前访问的配置相似，都会为访问的每个配置评估执行MD模拟所需的能量和力。这一建议的基本前提是，基于机器学习的方法可以用来消除在局部最小值和等价的多个局部最小值下预测重访或类似状态的原子力和能量所涉及的大量工作，从而消除大量冗余。只有当遇到真正新的构型时，才需要从头算起；否则，廉价的机器学习算法被用来预测原子力和能量。在前一种情况下，学习算法被重新训练以包括新的信息，从而使预测方案在运行中自适应。这项研究的具体目标是：(1)仅使用从头算MD模拟过程中积累的数据来发展独立的力场；以及(2)将这一发展应用于重要的表面科学模型问题，包括表面吸附原子扩散、表面氧化和表面催化反应。还计划了几项教育举措，包括开设新课程、讲习班、在线讲座、短期课程和座谈会。

1600218 PI: Ramprasad Institution: University of ConnecticutTitle: Accelerated Dynamics of Surface Chemical ReactionsSurface chemical reactions are ubiquitous in natural phenomena and play a key role in several scientific disciplines, such as heterogeneous catalysis, crystal growth, and electrochemistry. Achieving a molecular-level picture of the mechanisms and dynamical detail of surface chemical reactions though remains a daunting task. Current methods of inquiry, be it empirical or computational, provide us with only a limited view of this complex world. One way of achieving the requisite level of understanding is by the use of molecular dynamics (MD) simulations. These simulations can directly monitor the progression of reactions at surfaces at the molecular level as a function of a variety of relevant conditions. Nevertheless, significant gaps remain in present-day MD capabilities: they are either fast (but not versatile or accurate), e.g., those based on empirical force-fields, or they are versatile and accurate (but not efficient), e.g., those based on quantum mechanical (or ab initio) methods. The proposed work will exploit an adaptive machine learning scheme that can both accelerate ab initio MD simulations as well as create accurate force-fields (at no extra cost) on-the-fly. Timescales previously unreachable using quantum mechanical simulations may be accessed using this new paradigm (conceivably,milliseconds to seconds), while still preserving the fidelity of quantum mechanics.The primary reason ab-initio MD simulations are slow is largely because of enormous redundancies that permeate present-day paradigms. Energies and forces, the ingredients necessary to perform MD simulations, are evaluated for every configuration that is visited, regardless of whether a new configuration is similar to a previously visited configuration. The basic premise underlying this proposal is that a methodology based on machine learning can be used to eliminate the significant effort involved in predicting atomic forces and energies of revisited or similar states within a local minimum, and at equivalent multiple local minima, thus eliminating an enormous amount of redundancies. Only when a truly new configuration is encountered is an ab initio scheme necessary; otherwise, the inexpensive machine learning algorithm is used to predict atomic forces and energies. In the former instance, the learning algorithm is retrained to include the new information, thus making the prediction scheme adaptive on-the-fly. The specific goals of the proposed research are: (1) The development of stand-alone force-fields just using the data accumulated during ab initio MD simulations; and (2) Application of this development to important model surface science problems, including surface adatom diffusion, surface oxidation, and surface catalytic reactions. Several educational initiatives are also planned, including new course offerings, workshops, online lectures, short courses, and symposia.