Optimizing supercontinuum spectro-temporal properties by leveraging machine learning towards multi-photon microscopy

Abstract

Multi-photon microscopy has played a significant role in biological imaging since it allows to observe living tissues with improved penetration depth and excellent sectioning effect. Multi-photon microscopy relies on multi-photon absorption, enabling the use of different imaging modalities that strongly depends on the properties of the sample structure, the selected fluorophore and the excitation laser. However, versatile and tunable laser excitation for multi-photon absorption is still a challenge, limited by e.g. the narrow bandwidth of typical laser gain medium or by the tunability of wavelength conversion offered by optical parametric oscillators or amplifiers. As an alternative, supercontinuum generation can provide broadband excitations spanning from the ultra-violet to far infrared domains and integrating numerous fluorophore absorption peaks, in turn enabling different imaging modalities or potential multiplexed spectroscopy. Here, we report on the use of machine learning to optimize the spectro-temporal properties of supercontinuum generation in order to selectively enhance multi-photon excitation signals compatible with a variety of fluorophores (or modalities) for multi-photon microscopy. Specifically, we numerically explore how the use of reconfigurable (femtosecond) pulse patterns can be readily exploited to control the nonlinear propagation dynamics and associated spectral broadening occurring in a highly-nonlinear fiber. In this framework, we show that the use of multiple pulses to seed optical fiber propagation can trigger a variety of nonlinear interactions and complex propagation scenarios. This approach, exploiting the temporal dimension as an extended degree of freedom, is used to maximize typical multi-photon excitations at selected wavelengths, here obtained in a versatile and reconfigurable manner suitable for imaging applications. We expect these results to pave the way towards on-demand and real time supercontinuum shaping, with further multi-photon microscopy improvements in terms of spatial 3D resolution, optical toxicity, and wavelength selectivity.

Supercontinuum in integrated photonics: generation, applications, challenges, and perspectives

“…To that end, it is possible to use machine learning methods in order to have some additional control of the coherent dynamics. For example, genetic algorithms were used to prepare custom pulse-trains with an integrated pulse-splitter as to maximise the spectral density in defined wavelength ranges of the SC [ 236 , 237 ]. It was also shown that Gaussian-like peaks can be intentionally positioned in the spectrum of the generated SC, by changing the spectral phase of the incoming pulse with a genetic algorithm [ 238 ].…”

Structure optimization of large-solid-core photonic crystal fibers based on Ge\(_{20}\)Sb\(_{5}\)Se\(_{75}\) for optical applications

“…More recently, PCFs have transitioned from the confines of the laboratory to find utility in various fields such as optical communication, fiber light sources, fiber optic sensors, non-linear devices, and opto-fluidic devices [5][6][7]. Moreover, these fibers play a crucial role in the development of broad supercontinuum (SC) sources [8][9][10], which have found applications in a wide range of fields, including optical coherence tomography, optical metrology, and spectroscopy [11][12][13].…”

Conditional Recurrent Neural Networks for broad applications in nonlinear optics

“…It has been extensively applied in image recognition and classification [3,4], natural language processing [5], time-series prediction [6], cybersecurity [7], healthcare [8], autonomous vehicle control [9] and neuroscience research [10]. In the recent years, machine learning techniques have also been developed for different applications in optics [11][12][13], such as inverse design of photonic structures [14], and optical microscopy [15].…”