机器学习辅助非绝热分子动力学的可靠性验证

Reliability Validation for Machine-Learning-Assisted Nonadiabatic Molecular Dynamics

摘要: 机器学习能够降低非绝热分子动力学中重复电子结构计算的成本，其预测能力取决于模型能否保留驱动核运动与电子跃迁的多态关系。本文聚焦分子激发态面跳跃动力学，围绕表示选择、误差机制与验证证据开展方法综述。首先区分绝热性质学习与内部透热哈密顿量表示，再讨论多态信息共享、耦合相位处理、能隙加权及其与传播算法的连接。解析二态模型表明，能隙加权能够使耦合峰高保持有界，但峰宽仍随最小能隙减小而收缩；由加权目标恢复导数耦合时，能隙误差也会被放大。代表研究的比较进一步说明，数据划分应与泛化任务一致，动力学评价应同时报告布居定义、统计不确定性和失败轨迹。可靠的机器学习加速应以目标观测量为依据，联合检验电子结构、数据覆盖和传播近似，并在实际验证范围内解释其物理预测能力。

Abstract: Machine learning can reduce the cost of repeated electronic-structure calculations in nonadiabatic molecular dynamics, but predictive performance depends on retaining the multistate relationships that drive nuclear motion and electronic transitions. This review focuses on molecular excited-state surface-hopping dynamics and connects representation choices with error mechanisms and validation evidence. We distinguish direct learning of adiabatic properties from internal diabatic Hamiltonian representations, and examine multistate information sharing, phase treatment, energy-gap weighting, and their interfaces with propagation algorithms. An analytical two-state model shows that gap weighting bounds the coupling peak height, whereas the peak width still decreases with the minimum energy gap. Recovering derivative couplings from weighted targets also amplifies energy-gap errors. Comparisons of representative studies indicate that data splitting must match the intended generalization task, while dynamical assessments should specify population definitions, statistical uncertainty, and treatment of failed trajectories. Reliable acceleration should therefore be assessed through target observables, supported by joint tests of electronic properties, sampling coverage, and propagation approximations, with physical predictions restricted to the conditions actually validated.