密度泛函和神经网络组合高效热化学方法

密度泛函理论（DFT）已成为目前基于第一性原理处理复杂体系电子结构的最为有力的工具。但近年来不断有证据表明，以B3LYP为代表的DFT方法，存在着一些严重的缺陷，使得其应用于复杂体系的能力不断受到质疑。本项目拟以B3LYP 等DFT方法的计算结果为基础，系统地探讨分子组成与分子内不同结构和各种泛函预测误差之间的关系，结合人工神经网络与全局优化算法，构造合理的高效的计算模型，精确预测包括过渡金属化合物在内的标准生成焓、键能等热化学性质。提高DFT 应用于复杂体系的可靠性，并为新泛函的探索提供有价值的线索。

DFT has become a common tool for first principles quantum chemical computations of the electronic structure and properties. But it is now well-documented that B3LYP is not good for everything. The known problems include that it (1) always accumulates errors on heats of formation as the size of the system is increased (2) often fails to give reliable energy ordering of isomers (3) might underestimate reaction barrier heights (4) fails to give reliable description of vdW interactions, which limit its applications. In this project, we aim to set up a serie of combination methods with a neural network correction on top of the B3LYP or other functionals, in order to give heats of formation, bond energies, reaction barriers, isomerization energies in chemical accuracy. Furthermore, we will manage to alleviate most of the foregoing problems and extent its application on transition metal compounds.