神经网络全局势函数在多相催化中的应用

当今的多相催化研究需要新的技术和方法从原子尺度上表征活性中心结构和反应中间体。本文作者课题组近期开发了理论模拟新技术来探索催化剂活性位点结构，即基于神经网络势函数的大规模原子模拟（LASP）软件中实现的全局神经网络势函数计算方法。本文介绍了该方法可以显著降低催化体系的计算代价，而维持与密度泛函理论同一级别的计算精度，从而解决多相催化中的许多复杂问题。本文对神经网络势函数方法的实现细节和目前已实现的应用场景进行了详细介绍。神经网络势函数可以用来预测材料晶体结构，理解高压氢化条件下TiO₂表面的结构演化和确定三元氧化物ZnCrO晶相中合成气制甲醇活性位点。最后文章分析了神经网络势函数的局限性和今后可能的三个研究方向，即材料性质预测、多元素体系神经网络势函数构造和化学反应拟合。

Global neural network potential applications in heterogeneous catalysis

Heterogeneous catalysis demands new techniques and methods to characterize the structures of catalytic active centre and reaction intermediates from atomic scale. The recently developed global neural network potential (NNP) to explore catalyst structure is introduced, which has been implemented in the software of Large-scale Atomic Simulation with neural network Potential (LASP). The technical details of the NNP and its recent applications in heterogeneous catalysis are discussed. NNP can significantly reduce the calculation cost with comparable calculation accuracy to the ab-initio methods, with which many complex problems in heterogeneous catalysis can be solved. The success of NNP function in predicting the crystalline phase of materials, understanding the surface structure evolution of TiO₂ under high pressure of hydrogen and determining the active sites of ternary oxide are illustrated as examples. Finally, the limitations of NNP are discussed and its future research directions are pointed out as the estimation of material properties, the construction of NNP for multi-element system and the fitting of chemical reaction.