目前，为了精确控制光子-电子-声子耦合系统中的能量吸收、转换和耗散，吸附物-金属复合物必须表现出窄带局域表面等离子体共振（LSPR）、高能量效率和可调的阻尼系数。然而，由于实时时间依赖密度泛函理论（rt-TDDFT）的计算规模限制，解决涉及亚飞秒电荷相干性和纳米级几何结构的多尺度动力学问题仍面临困难，这阻碍了大规模非平衡载流子模拟的发展。本研究以一个 13 原子的吸附二氧化碳复合物作为测试平台。我们开发了一种改进的分层交互粒子神经网络（HIP-NN），并将其与矩传播理论（MPT）相结合。这使得能够构建一个受矩约束的分层交互粒子网络，该网络编码了电荷、二阶矩和距离矩阵。通过结合电荷/偶极子守恒和 MPT 损失，我们实现了从数小时缩短至 20 秒的 32 飞秒动力学推断，峰值误差通常小于 0.3 eV。在相同的 PBE/DZP rt-TDDFT 设置下，我们的模型在定量偏差很小的情况下重现了偶极子轨迹和吸收特征，同时提供了 540 倍的速度提升。该分析与金属调节的二氧化碳活化和热载流子机制一致，并可能有助于在亚飞秒至纳秒时间尺度上进行结构性能优化研究时的一致评估。该研究成果被著名期刊ACS Nano（IF=16.1,中科院一区TOP期刊）接受出版！出版中.....

全文链接：https://doi.org/10.1021/acsnano.5c14004

Figure 1. (a) CO₂–metal cluster system under a dipole-Gaussian pulse and the machine learning framework with an MPT branch. (b) Optimized geometries of CO₂ adsorbed on Ag₁₃, Pd₁₃, Rh₁₃, and Ru₁₃ clusters. (c) Data generation workflow from DFT simulations to the extraction of structural and electronic descriptors.

Figure 2 illustrates the Transition contribution map (TCM) combined with Density of States (DOS) overlays for the principal absorption peaks of CO₂/M₁₃ clusters containing Ag, Pd, Rh, and Ru. The peak energies are as follows: Ag₁₃ 3.27eV/4.12eV, Pd₁₃ 2.80eV/3.39eV, Rh₁₃ 5.54eV/6.21eV, Ru₁₃ 2.16eV/6.28eV.

Figure 3. Experimental calibration and method consistency settings for the absorption spectra of Ag_₁₁. Four normalized absorption spectra are presented in the figure: Ref. (FFT), MPT-ML, GPAW (rt-TDDFT), and Exp. (solid Ar matrix, 10 K, scanned from 2–6 eV). To ensure comparability, the three calculated curves and the experimental spectrum were processed with the same energy grid (2.5–5.5 eV) and the same broadening, and the intensities were normalized by the peak height. Based on this, the consistency of the main resonance energy and the spectral shape envelope was compared. (b) For the absorption spectrum of Ag_₅₅, TDDFTB (time-dependent DFT tight-binding) and TDDFT+TB were added outside the baseline in (a).