Development of a Neural Network Potential for Metal-Organic Frameworks

近年来，在开发用于原子模拟的机器学习(ML)潜力方面取得了许多进展。一类重要的ML势采用人工神经网络来构造原子构型与势能之间的函数关系。到目前为止，神经网络电位(NNP)已经被广泛地应用于各种材料中。他们被训练成来自电子结构计算的数据，然后允许以简单经验势的效率执行大型系统的模拟，同时保持基本参考方法的准确性。本项目将研究高维NNPs在有机-无机杂化材料中的适用性和准确性，这对传统的潜力是非常具有挑战性的。为此，金属有机骨架(MOF)将被用作一类典型的、具有重要技术意义的杂化材料。MOF由金属氧簇组成，它们通过有机连接物分子连接起来，形成非常稳定的多孔三维晶体材料。将特别注重对应适用于广泛的MOF的NNP的确认，这将对开发包含有机和无机子系统的一般混合系统的潜力产生影响。

A lot of progress has been made in recent years in the development of machine learning (ML) potentials for atomistic simulations. An important class of ML potentials employs artificial neural networks to construct the functional relation between the atomic configuration and the potential energy. To date, neural network potentials (NNP) have been reported for a wide range of materials. They are trained to data from electronic structure calculations and then allow to perform simulations of large systems with the efficiency of simple empirical potentials while maintaining the accuracy of the underlying reference method. In this project the applicability and accuracy of high-dimensional NNPs for organic-inorganic hybrid materials will be investigated, which are very challenging for conventional potentials. For this purpose metal-organic-frameworks (MOFs) will be used as a prototypical and technologically important class of hybrid materials. MOFs consist of metal-oxo clusters that are connected by organic linker molecules to form very stable porous three-dimensional crystalline materials. A particular focus will be on the validation of the NNP that should be applicable to a wide range of MOFs, with implications for the development of potentials for general hybrid systems containing organic and inorganic subsystems.