Efficient and scalable global structure optimization

分子簇不仅是连接单个分子和主体物质的纽带，而且在纳米技术、多相催化和气候研究等许多领域都有直接的意义。它们的结构和性质不同于单分子和块状物质，通常也不同于化学经验所预期的。因此，实用化的团簇结构预测具有重要意义。然而，这是很难实现的，因为簇结构的搜索空间随着簇的大小呈指数增长。因此，标准的局部优化工具严重不足，而且目前成熟的非确定性全局优化策略(NDGO；例如，进化算法(EA))对于较大的集群来说变得笨拙。然而，这正是星系团实验需要理论支持的地方。因此，本项目旨在提高全局聚类结构搜索效率，改善理论与实验之间的联系。本项目续篇的大部分内容都致力于EA方法的开发。对于局部优化，存在既通用又高效的算法。然而，“没有免费午餐”的定理已经证明，这种组合对于NDGO来说是不可能的。在这种情况下，必须通过利用特定于应用程序的功能来实现高效率。遵循这一核心思想，大多数开发线通过使EA适应手头的问题来追求更高的效率。然而，这需要的重要EA控制并不是强加给用户的；相反，EA被用来检测和“学习”必要的更改。因此，可以预期，由此产生的自适应进化算法不仅更高效，而且更易于操作。进化算法的另一部分开发和两个进化算法的应用关注于进化算法的一个新领域，即中等尺寸有机分子在表面的聚集。在第一个项目阶段，我们已经可以证明环境影响评估在这一领域是适用和有用的。然而，在上述意义上(通过特定问题的功能来提高效率)，还需要进一步的方法开发。这两个应用将与正在进行的“智能表面”和多相催化实验项目密切合作进行。

Molecular clusters are not only the link between single molecules and bulk matter but also have direct relevance in many areas, including nanotechnology, heterogeneous catalysis, and climate research. Their structures and properties differ from that of single molecules and bulk matter, and often also from what is expected by chemical experience. Therefore, practically applicable cluster structure prediction is of prime importance. However, this is hard to achieve since the search space of cluster structures grows exponentially with cluster size. Hence, standard local optimization tools are grossly insufficient, but also the by now well established non-deterministic global optimization strategies (NDGO; e.g., evolutionary algorithms (EA)) become unwieldy for larger clusters. Nevertheless, this is exactly where cluster experiments need theoretical support. Therefore, the present project aims at increases in global cluster structure search efficiency and at improved connections between theory and experiment.Large parts of this project continuation are devoted to EA method development. For local optimization, algorithms exist that are both general and highly efficient. However, the "no free lunch" theorem has proven that this combination is impossible for NDGOs. There, high efficiency has to be achieved by exploiting application-specific features. Following this core idea, most development lines aim at higher efficiency by adapting the EA to the problem at hand. However, the non-trivial EA control that this entails is not forced upon the user; instead, the EA is instrumented to detect and to "learn" the necessary changes. Hence, it can be expected that the resulting adaptive EA will not only be more efficient but also easier to handle.Another part of EA method development and two EA applications focus on an area that is new for EAs, namely aggregation of medium-size organic molecules on surfaces. In the first project phase, we could already demonstrate that EAs are applicable and useful in this area. In the above sense (efficiency via problem-specific features), however, further method development is needed. The two applications will be conducted in close collaboration with ongoing experimental projects, on "smart surfaces" and on heterogeneous catalysis.