激发态力场方法及其在非绝热动力学中的应用

Multireference approaches are very expensive to simulate photoinduced chemical and biochemical processes if not impossible. On the other hand, most of the existing molecular mechanics force field models with much higher computational efficiency are constructed based on the adiabatic potential energy surface in the electronic ground state. In the current proposal, we aim to develop a new force field method for excited states of molecules beyond the Born-Oppenheimer approximation using electronic structure calculations, mixed quantum-classical dynamics and machine learning. The primary goal of the proposed force field is to predict not only the potential energies in the ground and excited states but also the nonadiabatic couplings between different electronic states with high computational accuracy and efficiency simultaneously. The critical quantum effects on photochemical processes are expected to be reproduced with this model. We further plan to investigate the performance of molecular mechanics approaches such as the existing force fields on the multiscale quantum mechanics / molecular mechanics (QM/MM) calculations for excited states. We aim to replace the expensive multireference calculations within the framework of QM/MM by the suggested force field for excited states of molecules. The proposed work will lead to the advancement of theoretical and computational methodology both in the force field modelling and the excited state calculations. It will also serve to aid in the applications on diverse photochemical and photobiological systems in complex environments with long-time nonadiabatic molecular dynamics simulations.

如何准确高效地模拟和描述光诱导的化学和生物过程是当前理论及计算化学的一大难点。一方面，激发态量子力学方法的计算量巨大，只能处理相对简单的体系；另一方面，目前常用的力场方法大都局限于绝热势能面，主要适用于电子基态的热化学反应。针对上述问题，本项目将利用电子结构、非绝热动力学和机器学习等研究手段，结合基态分子力场和量子-经典混合方法，建立一个复杂体系多电子态耦合势能面的理论模型，发展一套超越绝热近似、适用于复杂体系多态耦合的激发态力场方法，既保证半经验量子化学或力场方法的计算效率，又达到多组态量子力学方法的计算精度，同时能够准确描述光化学过程中的非绝热效应等量子属性。本项目还将探讨分子力学方法在激发态计算中的适用性，把拟发展的激发态力场和现有基态分子力场有效结合起来，将其应用于溶液和蛋白等复杂环境中长时间尺度的光化学过程的非绝热动力学模拟，为解决光化学及光生物的实际问题提供新的理论工具。

如何准确高效地描述电子态之间的非绝热耦合是计算光化学的一大难点。虽然机器学习力场近年来已经取得了相当大的进展，并被国外课题组率先成功推广到激发态体系的非绝热动力学模拟，但在非绝热耦合矢量的数据集构建及其拟合结果的可迁移性等方面仍有缺陷。本项目针对深度学习力场的上述问题，对标力场的计算效率，发展了一个高效的、机器学习驱动的激发态动力学方法，将长短期记忆网络直接用于演化混合量子-经典动力学的电子运动方程，既能够准确描述非绝热效应，又免去了非绝热耦合矢量的计算和力场化，从而同步提高了激发态动力学的计算精度和计算效率，为光化学提供了新的理论工具。另一方面，复杂光化学及光生物体系的理论计算可靠与否，关键在于如何描述激发态势能面的关键区域，以及如何在经典动力学中处理量子效应。本项目针对上述问题，提出了一个描述多电子态势能面交叉区域的理论模型，探讨了数十种混合量子-经典动力学算法在激发态计算中的表现，进一步揭示了核自由度和电子自由度之间复杂多样的耦合机制，为评估理论方法在实际光化学问题中的适用性提供了重要依据。本项目还发展了适用于小样本光功能材料的、包含多尺度化学描述符的机器学习模型，发现了激发态分子结构与光催化反应动力学之间新的构效关系，为进一步发展和改进激发态力场储备了关键技术和数据。