Lin Qidong, Zhang Liang, Zhang Yaolong, Jiang Bin
Hefei National Laboratory for Physical Science at the Microscale, Department of Chemical Physics, Key Laboratory of Surface and Interface Chemistry and Energy Catalysis of Anhui Higher Education Institutes, University of Science and Technology of China, Hefei, Anhui 230026, China.
J Chem Theory Comput. 2021 May 11;17(5):2691-2701. doi: 10.1021/acs.jctc.1c00166. Epub 2021 Apr 27.
Neural network (NN) potential energy surfaces (PESs) have been widely used in atomistic simulations with ab initio accuracy. While constructing NN PESs, their training data points are often sampled by molecular dynamics trajectories. This strategy can be however inefficient for reactive systems involving rare events. Here, we develop an uncertainty-driven active learning strategy to automatically and efficiently generate high-dimensional NN-based reactive potentials, taking a gas-surface reaction as an example. The difference between two independent NN models is used as a simple and differentiable uncertainty metric, allowing us to quickly search in the uncertainty space and place new samples at which the PES is less reliable. By interfacing this algorithm with the first-principles simulation package, we demonstrate that a globally accurate NN potential of the H + Ag(111) system can be constructed with merely ∼150 data points. This PES can be further refined to describe H dissociation on Ag(100) by adding ∼130 more configurations on this facet. The entire process is completely automatic and self-terminated once the relative error criterion is fulfilled. Impressively, data points sampled by this uncertainty-driven strategy are substantially fewer than by the traditional trajectory-based sampling. The final NN PES not only converges well the quantum dissociation probability of the molecule but also well-reproduces the phonon properties of the substrate and is capable of describing surface temperature effects. These results show the potential of this active learning approach in developing high-dimensional NN reactive potentials in gas and condensed phases.
神经网络（NN）势能面（PESs）已被广泛应用于具有从头算精度的原子模拟中。在构建NN势能面时，其训练数据点通常通过分子动力学轨迹进行采样。然而，对于涉及罕见事件的反应系统，这种策略可能效率低下。在这里，我们以气体 - 表面反应为例，开发了一种不确定性驱动的主动学习策略，以自动且高效地生成基于高维NN的反应势能。两个独立NN模型之间的差异被用作一种简单且可微的不确定性度量，使我们能够在不确定性空间中快速搜索，并在PES不太可靠的位置放置新样本。通过将该算法与第一性原理模拟软件包相结合，我们证明仅用约150个数据点就可以构建H + Ag(111)系统的全局精确NN势能。通过在该晶面上添加约130个更多构型，可以进一步细化该PES以描述H在Ag(100)上的解离。一旦满足相对误差标准，整个过程将完全自动且自终止。令人印象深刻的是，由这种不确定性驱动策略采样的数据点比传统的基于轨迹的采样要少得多。最终的NN势能面不仅很好地收敛了分子的量子解离概率，而且很好地再现了衬底的声子特性，并且能够描述表面温度效应。这些结果表明了这种主动学习方法在开发气相和凝聚相中高维NN反应势能方面的潜力。
