【摘要】 高次谐波产生(HHG)是强场物理和阿秒科学中的重要非线性过程,但在复杂介质和宽参数空间下,其数值模拟与参数优化面临较高的计算成本。本文结合本课题组的相关研究工作,综述了人工智能(AI)在高次谐波模拟与预测中的应用进展,重点介绍了代理模型、时间序列预测、贝叶斯优化及逆问题求解等方法在提高计算效率和物理信息提取能力方面的作用,并对AI在该领域的新兴方向和未来发展进行了讨论。更多还原

【Abstract】 Significance High-order harmonic generation(HHG) is an extreme nonlinear optical phenomenon arising from the interaction of intense laser fields with matter. Owing to its high photon energy, attosecond pulse capability, broad spectral coverage, and excellent spatiotemporal coherence, HHG has become the primary means of generating coherent extreme-ultraviolet to soft X-ray radiation. The attosecond pulses produced via HHG provide a powerful tool for real-time observation of electron dynamics, driving advances across fields from quantum control to optoelectronic devices. Despite its significance, HHG faces substantial computational challenges across different media. In gas-phase HHG, the interplay between microscopic responses and macroscopic propagation greatly increases system complexity and computational cost. In solid-state HHG, the full semiconductor Bloch equations(SBE) are expensive to solve, and key physical parameters such as dephasing times and band structures are difficult to measure directly. Plasma HHG typically requires particle-in-cell simulations, which are also computationally demanding. These challenges have motivated the integration of artificial intelligence(AI) into HHG research.Progress Recent advances in AI-assisted HHG research fall into three categories: forward prediction acceleration, inverse problem solving, and physically constrained interpretable modeling. In forward prediction, surrogate models have emerged as effective substitutes for costly quantum simulations by learning the mapping from laser parameters to microscopic dipole responses, achieving high accuracy at reduced computational cost. Time-series prediction methods reconstruct dipole moment evolution, enabling accurate harmonic spectra prediction under limited data conditions. Bayesian optimization and related algorithms further support automated parameter tuning for experimental control.In inverse problem solving, the focus shifts from fast computation to extracting hidden structural information from harmonic spectra. AI methods have been applied to retrieve driving field parameters, material properties, dynamical parameters, and single-molecule information directly from solid-state HHG spectra, bypassing experimental and computational barriers that traditional approaches face.To address the limited generalizability and interpretability of existing models, emerging physically constrained architectures have been developed along three lines. First, hidden physical features are uncovered: Zhao et al. identified non-adiabatic band-coupling features inaccessible to standard Fourier analysis. Second, physical laws are embedded as model constraints: Sen et al. proposed the Kolmogorov-Arnold network-Ehrenfest time-series analysis(KAN-ETS) framework, incorporating the Kolmogorov-Arnold representation theorem, Ehrenfest theorem, and temporal causality to enforce physical consistency while enhancing feature extraction. Third, wavefunction-level modeling is pursued: Huang et al. parameterized the time-evolving wavefunction using radial basis function(RBF) neural networks within a stochastic representation framework.Overall, AI applications in HHG have evolved from computational acceleration and spectral reconstruction, to parameter extraction, and further toward physically constrained and interpretable modeling. However, the generalizability and interpretability of current models remain underexplored, leaving substantial room for future investigation.Conclusions and Prospects AI has become a key enabler in HHG research, demonstrating clear advantages in computational acceleration, parameter retrieval, and physical feature extraction. However, several challenges persist, including the scarcity and quality of training data, the early-stage development of physically constrained and interpretable architectures, insufficient experimental validation and closed-loop feedback, and limited computational resources. Looking ahead, several directions merit attention: 1) further development of time-series modeling approaches; 2) systematic embedding of physical constraints to improve generalizability and interpretability, alongside AI-driven discovery of hidden physical features; 3) broader application of transfer learning; 4) AI-assisted modeling for emerging media such as liquids and plasma mirrors; and 5) application of generative AI and large models to high-throughput HHG experimental data. As AI and strong-field physics converge more deeply, HHG research holds the potential to transition from computation-driven toward a new paradigm of data-physics co-driven discovery.更多还原