Global optimization of reactive force fields

如果通过经典力学计算原子运动，并且从简单函数（力场）中提取原子之间的力，那么由数十亿原子组成的大分子组合的动力学可以在微观上长时间（高达毫秒）进行模拟。特别是在生物化学领域，在过去的几十年里建立了几个标准力场，允许模拟具有明确溶剂化的完整蛋白质，但不允许化学反应（破坏或形成化学键）。这些力场的使用是如此普遍，以至于一些专家认为力场模拟真的无法捕捉化学反应。然而，原则上这是可能的。实际上，在化学文献中有几个反应力场，但它们并不为人所知。一个主要原因是，无功力场比无功力场更复杂，包含更多的参数。成功地使用力场需要将其参数拟合到参考数据中。大量的参数将这一拟合转化为一个高度复杂的优化问题，其中有许多质量差异很大的最优解。因此，应用传统的优化方法需要大量的工作和经验，只是为了将一个坏的参数集改进到一个不太坏的参数集。然而，现代全局优化方法非常适合这样的任务。因此，在这个项目中，我们将利用我们多年来在遗传/进化算法方面的经验，并将其应用于反作用力场的全局优化。我们对测试用例的初步研究表明这是可能的并且是成功的。此外，他们还发现了几个层次上令人尴尬的并行化的可能性。因此，这种优化计算可以从现代的大规模并行计算机体系结构中获益良多。在这个项目中，我们将研究遗传/进化算法对反应力场全局参数优化的现实和一般适用性，使用不同层次的参考数据（直到高端电子结构结果），用于几种不同的反应力场和几种不同的真实化学体系。我们计划将这些研究产生的优化例程分发给用户社区，既可以作为灵活的独立应用程序，也可以嵌入到标准量子化学包中。这将允许最终用户生成自己的反应力场，特别是他们选择的化学反应，因此具有更高的精度，并将其应用于大规模的反应性分子动力学模拟。

The dynamics of large molecular assemblies, comprising billions of atoms, can be simulated for microscopically long times (up to the regime of milliseconds), if atomic movement is calculated via classical mechanics and if forces between atoms are extracted from simple functions (force fields). Particularly in the area of biochemistry, several standard force fields were established in the past decades, allowing for simulations of complete proteins with explicit solvation, but not allowing for chemical reactions (breaking or forming chemical bonds). Usage of these force fields is so commonplace that even some experts assume that force-field simulations really cannot capture chemical reactions. In principle, however, this is possible. Actually, there are several reactive force fields in the chemical literature, but they are not well known. A main reason for this is that reactive force fields are more complex and contain far more parameters than non-reactive ones. Successful use a force field requires fitting of its parameters to reference data. A large number of parameters transform this fitting into a highly complicated optimization problem with very many optima of widely varying quality. Application of traditional optimization methods hence requires a lot of work and experience, just to improve a bad parameter set to a less bad one.Modern global optimization methods, however, are very well suited for such tasks. Hence, in this project, we will draw upon our long years of experience with genetic/evolutionary algorithms and apply them to the global optimization of reactive force fields. Our preliminary studies on test cases have shown that this is possible and successful. In addition, they have uncovered possibilities for embarrassing parallelization on several levels. Hence, such optimization calculations can be expected to profit significantly from modern, massively parallel computer architectures.In this project, we will investigate the realistic and general applicability of genetic/evolutionary algorithms to global parameter optimization of reactive force fields, using reference data on different levels (up to high-end electronic structure results), for several different reactive force fields and for several different real chemical systems. We plan to distribute the optimization routines resulting from these studies to the user community, both as flexible stand-alone application and embedded into standard quantum-chemistry packages. This will allow end-users to generate their own reactive force fields, specifically for chemical reactions of their choice and hence with heightened accuracy, and to apply them in large-scale reactive molecular dynamics simulations.