Beyond harmonic transition state theory for accelerating molecular dynamics

来自德克萨斯大学奥斯汀分校的Graeme Henkelman在化学理论、模型和计算方法（CTMC）项目和网络基础设施办公室（OCI）的支持下，开发了加速分子动力学模拟的算法。对于罕见事件系统，如固体中的扩散和表面的反应，过渡态理论（TST）允许在分子振动和感兴趣的反应事件之间的时间尺度分离。当反应机理未知时，采用自适应的动力学蒙特卡罗（KMC）方法寻找鞍点，并用谐波TST评估速率。为了减少计算成本，使用数据库来存储动力学事件，并使它们可供以后在具有类似化学性质的单独计算中使用。当一类系统的反应机制已知时，计算成本很低，因为所有事件都是从数据库中提取的。为了进一步加速模拟，使用主方程的解析解来逃避由快速速率连接的状态组。存储这些状态的本地连通性对于提高效率至关重要。精度将提高与一组方法，从谐波TST到（经典）精确的全TST加上动态修正。TST的挑战是找到一个分离表面，将反应物与任何产物状态分开。目前正在开发一种支持向量机，以提供基于学习数据的分析分类功能。决策表面周围的采样点将用于训练机器，以便它可以在没有事先了解反应机制的情况下提供准确的过渡状态。这项工作旨在减轻模拟现实化学或材料过程所需的超长计算时间，例如催化、蛋白质折叠、分子扩散等。基本分子运动的特征时间比我们在实验室测量的时间尺度快很多很多倍，因此必须开发加速算法。EON2是由PI和合作者开发的一个分布式开源程序，用于计算系统中的长时间分子动力学，例如，在金属基质上进行催化和聚集。它将使用先进的数据库技术和机器学习，并将通过定制界面与其他软件结合使用。PI为任何对算法、软件或科学有疑问的人开设了一个讨论论坛。这个项目已经开放，这样校园里的计算机，空闲时间的超级计算机，以及公众中的任何人都可以贡献计算资源来加速原子尺度上的动力学模拟。外联计划的这一方面使公众能够直接访问作为该项目的一部分正在进行的研究。

Graeme Henkelman from the University of Texas at Austin is supported by the Chemical Theory, Models and Computational Methods (CTMC) Program and the Office of Cyber Infrastructure (OCI) in developing algorithms for accelerating molecular dynamics simulations. For rare event systems, such as diffusion in solids and reactions at surfaces, transition state theory (TST) allows for a separation of time scales between that of molecular vibrations and the reactive events of interest. When reactive mechanisms are not known, an adaptive version of the kinetic Monte Carlo (KMC) method is used to find saddle points and rates are evaluated with harmonic TST. To reduce the computational cost, a database is employed to store kinetic events and to make them available for later use in separate calculations with similar chemistry. When the reaction mechanisms for a class of system are known, the computational cost is low because all events are drawn from the database. To further accelerate the simulations, groups of states connected by fast rates are escaped using an analytic solution to the master equation. Storing the local connectivity of these states is essential for efficiency. Accuracy will be improved with a set of methods ranging from harmonic TST to the (classically) exact full TST plus dynamical corrections. The challenge of TST is finding a dividing surface that separates reactants from any product state. A support vector machine is being developed to provide an analytic classification function based upon learned data. Sampled points around the decision surface will be used to train the machine so that it can provide an accurate transition state without prior knowledge of reaction mechanisms.This work is directed at alleviating the ultralong computational times needed for simulation of realistic chemical or material processes, e.g., catalysis, protein folding, molecular diffusion and so on. The characteristic times for basic molecular motion are many, many times faster than the time scales we measure in the laboratory, and so acceleration algorithms must be developed. EON2 is being developed as a distributed open-source program developed by the PI and collaborators to calculate long-timescale molecular dynamics in systems, for example, undergoing catalys and clustering on metal substrates. It will use advanced database techniques and machine learning, and will be usable in conjunction with other software through customized interfaces. A discussion forum is hosted by the PI for anyone with questions about the algorithms, software or science. The project has been opened up so that computers on campus, supercomputers with idle time, and anyone from the public can contribute computational resources to accelerate dynamics simulations at the atomic scale. This aspect of the outreach program gives the public direct access to the research being done as part of this project.