Development of a generally applicable machine learning potential with accurate long-range electrostatic interactions

近年来，基于机器学习技术的新一代原子间势被引入。这些电势提供了原子位置与势能之间的直接函数关系，将电子结构方法的准确性与简单经验电势的效率结合起来。由于缺少特定于系统的术语，它们允许对各种各样的系统进行扩展模拟。这些电势大多依赖于原子的性质，比如能量和电荷，而这些性质只取决于原子的局部化学环境。然而，这种局部电荷无法捕获远程电荷转移。这阻碍了系统的准确描述，其中远距离结构特征对系统中的电荷分布具有全局影响。这类系统的例子有半导体，包括缺陷、氧化物的极性表面和具有不同可能金属氧化态的金属有机分子。为了克服当前机器学习潜力的这些固有局限性，我们提出将高维神经网络与电荷平衡神经网络技术相结合。由此产生的新方法将普遍适用于所有类型的系统，我们将通过使用最小跳变方法分析涵盖所有类型键的不同模型系统的势能面来证明这一点。

In recent years, a new generation of interatomic potentials based on machine learning techniques has been introduced. These potentials, which provide a direct functional relation between the atomic positions and the potential-energy, combine the accuracy of electronic structure methods with the efficiency of simple empirical potentials. Because of the absence of system-specific terms they allow to perform extended simulations of a large variety of systems. Most of these potentials rely on atomic properties like energies and charges depending only on the local chemical environments of the atoms. Such local charges are, however, unable to capture long-range charge transfer. This prevents the accurate description of systems in which distant structural features have global effects on the charge distribution in the system. Examples for such systems are semiconductors including defects, polar surfaces of oxides and metal-organic molecules with different possible metal oxidation states. In order to overcome these intrinsic limitations of current machine learning potentials, we propose to combine high-dimensional neural networks with the charge equilibration neural network technique. The resulting new method will be generally applicable to all types of systems, which we will demonstrate by analyzing the potential-energy surfaces of different model systems covering all types of bonding using the minima hopping method.