Machine learning for probabilistic inverse design in nanophotonics

• Autores: Michel Frising
• Directores de la Tesis: Ferry Prins (dir. tes.), Jorge Bravo Abad (dir. tes.)
• Lectura: En la Universidad Autónoma de Madrid ( España ) en 2022
• Idioma: inglés
• Número de páginas: 114
• Títulos paralelos:
  • Aprendizaje automático para diseño inverso probabilístico en nanofotónica
• Enlaces
  • Tesis en acceso abierto en: Biblos-e Archivo
• Resumen

  • Machine Learning has found its way into every aspect of our life from search algorithms, to virtual assistants to enhancing photos when taking images with a smartphone. Not surprisingly, the tremendous promise and potential of machine learning has been discovered by researchers and engineers working in photonics. Tailoring optical devices to have a specific response is a long standing and extremely hard problem, especially since the design space tends to be very high-dimensional. Despite the impressive results achieved so far, current inverse design, the task of reverse-engineering or finding a structure that exhibits a certain desired behavior, requires a lot of domain expertise and physical intuition. Generative models based on machine learning and ideas from Bayesian Inference and Optimization can usher in a new era of data-driven inverse design in photonics where the human is completely taken out of the loop. These models can be completely agnostic of the physics governing the response of a device or they include information from physical models to speed the design process. This thesis aims to connect the field of Nanophotonics with the powerful generative models that have been developed in the last years in the machine learning community. I will discuss how machine learning can learn to approximate complex probability distributions which can be used to describe distributions of devices exhibiting the same behavior. I demonstrate the power of these models by solving problems with complex multimodal loss functions which are typically hard to optimize with traditional methods. Further we investigate two physical systems that are conceptually simple, a slit in a thin metal film flanked by periodic corrugations and a stratified dielectric medium, to benchmark new neural network architectures to solve inverse design problems in nanophotonics. We finish that discussion by presenting a workflow to find out when generative methods are necessary and when not. The power of these models is demonstrated by performing inverse design on a layered structure for thermophotovoltaic applications and which is identical to a struciii ture reported from literature obtained with an evolutionary algorithm. Finally, a setup is described to automatically acquire polarization and angle-resolved spectra of ten-thousands of photonic structures to build a completely experimental dataset for machine learning to automatically design plasmonic filters based on periodic arrangements of sub-wavelength holes in metal films