Modeling of light-matter interactions with neural networks

Abstract

With the invention of the laser, the chemists' dream to investigate, initiate and control chemical reactions on the molecular scale seemed to come within grasp. However, the true beginnings of the realization of this dream had to wait until the time when the boundaries on the duration of laser pulses were pushed into the picosecond to femtosecond regime, the natural timescale of nuclear motions, i.e., vibrations and rotations in molecules. While such light sources allowed, for the first time, the direct observation of molecular dynamics, the breakthrough for the control of physical or chemical processes on the atomic or molecular scale came with the advent of pulse shaping techniques that made it possible to modify these ultrashort pulses, and hence to adapt them to the involved dynamics.

Applying artificial neural networks to coherent control experiments: A theoretical proof of concept

“…Because the literature related to this topic is already quite extensive (see, e.g., Refs. [18][19][20][31][32][33]), we will here only provide a brief general description of ANNs and their operating principles.…”

“…Furthermore, we suggest that an ANN that has been trained in this manner may be able to achieve the same control objective in a real experimental situation. Note that while ANNs have been used in the past to generate predictive models of ultrafast laser-molecule interactions [19,20], to our knowledge they have not been applied to coherent control experiments before.…”

Deep Learning Meets Nanophotonics: A Generalized Accurate Predictor for Near Fields and Far Fields of Arbitrary 3D Nanostructures

“…Examples are the possibility of phase recovery in conventional microscopy, stabilization of lasers, and a large variety of applications in data analysis and interpretation. − For forward modeling, early work has shown that artificial neural networks (ANNs) can be used as approximate predictors for light–matter interaction phenomena or optical scattering at nanostructures. Examples are strong-field ionization of potassium atoms, SHG of a specific fluorescent molecule, the optical transmittance of “H”-shaped particles or the scattering cross-sections of multilayer spheres . All of those reported ANN techniques apply to single, very specific problems and for a particular nanostructure geometry.…”

Deep learning in nano-photonics: inverse design and beyond

“…Early works have proposed ANNs to create phenomenological models of non-linear optical effects or of (iv) shows a numerical simulation of the field intensity to test the ANN-design. [50] optical ionization using experimental training data [51,52]. Recently, the idea has been picked up and it has been shown for instance that scattering and extinction spectra can be predicted with high accuracy [53] and also that the phase can be included in the predictions [54], which is important for nanostructures in metasurfaces.…”