【摘要】 超快激光与物质的相互作用过程通常伴随强非线性效应与高维参数耦合,导致其表征与调控面临诸多困难。传统依赖精确物理模型、迭代反演或经验调参的方法在实时性、鲁棒性及泛化能力等方面存在局限性。近年来,人工智能(AI)凭借在数据分析速度与高维参数空间建模等方面的显著优势,为超快脉冲信息获取与光场调控提供了新的技术范式。本文首先系统介绍深度学习与强化学习等典型AI方法及其在超快光学中的应用潜力,进而从“数据驱动的逆问题求解”与“学习型闭环控制”2条主线出发,综述AI在超快脉冲表征、时域脉冲整形、空间光调制、强非线性传输及超连续谱整形等方向的代表性研究进展。在此基础上,重点探讨当前AI方法在跨系统泛化、物理先验融合、实时闭环调控和多目标长期稳定运行等方面面临的关键挑战。未来,AI有望推动超快脉冲表征与光场调控迈向更高精度、更强鲁棒性及可长期稳定运行的工程化应用阶段。更多还原

【Abstract】 Significance The precise characterization of ultrafast pulses and the ability to control optical fields are central challenges in modern photonics. These capabilities underpin a wide range of applications, including high-speed optical communication, ultrafast spectroscopy, attosecond science, and nonlinear optical phenomena. Traditional methods often struggle with limitations such as measurement constraints, strong system nonlinearities, and high-dimensional parameter spaces, which restrict the achievable precision and efficiency in experiments. Advancements in this area are therefore critical for pushing the frontiers of ultrafast optics and enabling new scientific discoveries.The introduction of artificial intelligence into ultrafast pulse characterization and light field manipulation offers a transformative paradigm. Artificial intelligence(AI)-assisted approaches, such as deep learning and reinforcement learning, can extract pulse information from limited measurements, optimize complex nonlinear systems, and implement real-time, adaptive control. Developing these methods is essential for achieving higher precision, enhanced robustness, and long-term stable operation, paving the way for next-generation optical technologies and engineering applications that have been previously unattainable.Progress AI-assisted ultrafast pulse characterization refers to a method that utilizes artificial intelligence to reconstruct the electric field(amplitude and phase) of an ultrashort pulse from limited experimental measurements. By treating pulse characterization as a nonlinear inverse problem, neural networks learn the complex mapping from measured data-such as frequency-resolved optical gating(FROG) traces, interferometric correlation traces, or spatial intensity patterns-directly to the pulse parameters. The advantage of this AI-driven approach lies in its ability to perform reconstructions in illiseconds with high noise robustness, reducing reliance on iterative algorithms, precise system models, and initial guesses. It can be integrated with traditional methods like FROG to accelerate reconstruction or combined with novel physical encoding schemes, such as using a multimode fiber to encode temporal information into a single-shot spatial image, thereby enabling real-time, single-shot characterization. AI-assisted ultrafast light field manipulation is a method that leverages AI to manage the high-dimensional and strongly nonlinear nature of optical systems, enabling efficient and robust temporal pulse shaping, wavefront correction, and nonlinear spectral shaping. In temporal pulse shaping, AI algorithms learn the system’s response to control parameters [e.g., from a deformable mirror(DM) or spatial light modulator(SLM)] to accelerate the closed-loop optimization process, converging on target pulse shapes with significantly fewer iterations than traditional evolutionary algorithms. In spatial light modulation, AI enables rapid generation and control of complex spatial phase distributions; for instance, deep learning can directly predict digital micromirror device(DMD) or SLM control patterns from target wavefronts to implement continuous phase modulation in thermo-optically addressed liquid-crystal devices or achieve high-fidelity multi-wavelength focusing through multimode fibers by learning the mapping between modulation masks and output intensity patterns. Finally, for strong nonlinear propagation, deep reinforcement learning enables online adaptive control to stabilize supercontinuum generation, predicts control parameters for target spectral shapes, and creates surrogate models to accelerate multi-objective optimization in nonlinear processes.Conclusions and Prospects This article provides a comprehensive review of the applications of AI in ultrafast pulse characterization and optical field control. These techniques significantly enhance the accuracy and efficiency of pulse measurements and the capability of optical field manipulation by learning the complex mapping from limited experimental measurements to pulse parameters, or by adaptively optimizing control strategies in high-dimensional nonlinear systems. Studies have shown that in pulse characterization, AI can achieve rapid, high-signal-to-noise single-shot pulse reconstruction by integrating nonlinear and linear measurement approaches. In optical field control, AI demonstrates notable advantages in temporal pulse shaping, spatial light modulation, and the control of strongly nonlinear propagation processes, such as supercontinuum generation, greatly improving optimization efficiency and system robustness. Despite these advances, current methods still face challenges including insufficient physical consistency and interpretability of models, difficulties in cross-system generalization, and long-term stability in controlling strongly nonlinear processes. Therefore, deeply integrating physical priors and developing algorithms with stronger generalization capability and uncertainty estimation remain key research directions. Furthermore, to achieve complex optical field control objectives under real-time and long-term stable operation, it is necessary to further explore more reliable closed-loop coupling between learned surrogate models and online control strategies. The engineering applications of AI in ultrafast optics are expected to provide new impetus for the advancement of ultrafast imaging, precision manufacturing, and strong-field physics.更多还原