机器学习策略下的超快光子学设计:回顾与展望(特邀)

Intelligent Ultrafast Photonics Based on Machine Learning:Review and Prospect(Invited)

从行为组织学开创了光子计算的先河以来,基于人工智能的光学计算已经发展了七十多年,这一历程对超快光子学的智能化研究产生了重要影响。近年来,因超短脉冲非线性多维相互作用的复杂化,让超快光子学方向的研究产生了巨大的发展潜力。智能超快光子学的研究,为超短脉冲数据的完整、准确和有代表性提供了新的推动力量。在这里,我们回顾了机器学习策略下超短脉冲光纤激光系统的最新进展。通过算法和控制元件两方面的设计,进一步概述了满足这些进展所需的技术条件。并对机器学习与超快光子学这一新兴交叉技术所存在的挑战与未来研究前景做出展望。

Artificial intelligence-based optical computing has evolved for over seventy years since behavioral histology pioneered photonic computing,a journey that has had a significant impact on the study of intelligence in ultrafast photonics.The use of machine learning techniques in ultrafast photonics is an exciting new field,positioning innovative research in fundamental science and cutting-edge technologies.On the one hand,the systems of ultrafast photonics and the techniques that generate information reproduction,transformation and transmission can be used for photonics implementation of machine learning techniques;on the other hand,the powerful and precise,efficient and flexible data analysis and processing capabilities of machine learning can solve the challenges of ultrafast photonics in design and control.In recent years,the development of research in ultrafast photonics has been greatly hampered by the complexity of nonlinear multidimensional interactions of ultrashort pulses.The research of smart ultrafast photonics provides a new driving force for complete,accurate and representative data of ultrashort pulses.The prospect of machine learning in generating ultrashort pulse lasers has been realized by the intelligent design and operation of ultrafast fiber lasers composed of nonlinear photonic devices based on saturable absorbers to control photonic elements to produce nonlinear effects.At the same time,machine learning strategies are used to optimize control algorithms and feedback loops to achieve technological breakthroughs in nanophotonics and pulse dynamics.This results in the characterization and control of ultrafast photonics.Here,we review recent advances in intelligent machine learning-based ultrashort pulsed fiber laser systems,further outlining the scientific and technical conditions required to meet these advances through algorithmic foundations and key architectures.We also provide an outlook on the challenges and future research prospects for the emerging cross-technology direction of machine learning and ultrafast photonics.The development of ultrafast photonics with machine learning strategies has great potential.On the one hand,in terms of hardware,the intelligence of the modules of the laser system is an important development direction.So far,optical technology has been mainly optimized in a limited design space,limiting the photonic structure.Further development of advanced hybrid schemes will be a key step to solving this challenge,and the combination of smart optoelectronic devices will greatly contribute to the advanced level of lasers and expand the design space of photons to achieve the best performance of the target system.On the other hand,from an algorithmic point of view.To control more dimensional variables in the feedback,the algorithm′s efficiency becomes crucial.So far,most designs are based on genetic algorithms or neural network architectures.While these implementations have undoubtedly led to remarkable and pioneering results,the fact remains that a combination of strategies may be required to exploit the full potential of machine learning.Revealing models from experimental data,forming neural networks after countless data mining sessions,and establishing infinite training relationships between targets and parameters will greatly optimize the time-consuming performance of ultrafast lasers to quickly locate the desired operating states.In addition,unsupervised learning algorithms,including clustering and expectation-maximization,have the ability to infer and reveal hidden internal structures from data without labeled responses,which may have a significant role in key problems such as dimensionality reduction of complex nonlinear systems.Similarly,real-time adjustment of parameters by monitoring the collection of various information about the environment serves as a regularizer for deep learning models to ensure that optical research relies on stability and longevity.In the future,improving the portability of intelligent photonic systems and designing inference models as specific products are important topics for exploring the next generation of ultrafast photonic technologies.The convergence of machine learning and ultrafast photonics cutting-edge crossover technologies in the context of artificial intelligence takes an unconventional approach to provide an unparalleled photonic perspective.This intersection of computer science,photonics,and materials platforms will enable new approaches to the large-scale photonic design of unique functions as well as optical characterization,laying the cornerstone for efficient energy conversion systems.We envision that a global optimization framework based on a multi-step machine learning strategy can build a more general intelligent ultrafast photonic system,where the first step can be to define the main target function of the device and determine the appropriate photonic concept to provide the best performance.The second step is to select a suitable material platform and build an extensive database of optical materials.By using the selected material properties,an optimized design solution for the material device can be provided.The third step is to determine the appropriate fabrication conditions(growth conditions,doping levels,stoichiometry,etc.)and integration schemes.The interplay between new photonic structures and machine learning may overcome the limitations of current computational methods and systems,provide unparalleled capabilities in light-matter interactions and unlock new device concepts,and may lead ultrafast photonics research to new frontiers that could usher in a brighter era of artificial intelligence.