Department of Physics and Astronomy, and Interdisciplinary Nanoscience Center (iNANO), Aarhus University, DK-8000 Aarhus C, Denmark.
J Chem Phys. 2018 Jun 28;148(24):241734. doi: 10.1063/1.5023671.
We present a method for accelerating the global structure optimization of atomic compounds. The method is demonstrated to speed up the finding of the anatase TiO(001)-(1 × 4) surface reconstruction within a density functional tight-binding theory framework using an evolutionary algorithm. As a key element of the method, we use unsupervised machine learning techniques to categorize atoms present in a diverse set of partially disordered surface structures into clusters of atoms having similar local atomic environments. Analysis of more than 1000 different structures shows that the total energy of the structures correlates with the summed distances of the atomic environments to their respective cluster centers in feature space, where the sum runs over all atoms in each structure. Our method is formulated as a gradient based minimization of this summed cluster distance for a given structure and alternates with a standard gradient based energy minimization. While the latter minimization ensures local relaxation within a given energy basin, the former enables escapes from meta-stable basins and hence increases the overall performance of the global optimization.
我们提出了一种加速原子化合物全局结构优化的方法。该方法在密度泛函紧束缚理论框架下，利用进化算法，成功加快了锐钛矿 TiO(001)-(1×4)表面重构的发现速度。作为该方法的关键要素，我们使用无监督机器学习技术将多样化的部分无序表面结构中的原子划分为具有相似局部原子环境的原子簇。对 1000 多个不同结构的分析表明，结构的总能量与特征空间中各原子环境与其各自簇中心之间的距离总和相关，其中总和遍历每个结构中的所有原子。我们的方法被表述为给定结构中此总和簇距离的基于梯度的最小化，并与基于梯度的标准能量最小化交替进行。后者的最小化确保了给定能量盆地内的局部松弛，而前者则实现了从亚稳盆地的逃逸，从而提高了全局优化的整体性能。
