AI / ML for photonics
Data-driven methods of machine learning (ML) have attracted a lot of interest in various fields of physics. Inverse design and optimisation of structured optical metamaterials such as photonic crystals, metasurfaces, and other nanostructured components seem to benefit a lot from this approach in the nearest future. We develop several approaches to use ML methods to effectively predict and optimise properties of photonic crystals and metamaterials (e.g. size of photonic bandgaps, effective medium parameters, photonic band diagrams). For inverse design, we use generative ML models such as Variational Auto-Encoders (VAEs), Generative Adversarial Networks (GANs), and also are investigating other newest developments in this field.
Our recent research is outlined in special issue on Inverse Design of nanophotonics devices and materials https://authors.elsevier.com/a/1ftCg5asNsdBxW (see also open-access early version of this work https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4204785 ) and future projects are related to 3D extensions of this research and applications of various ML methods to the data from the newest databases of photonic crystals (see e.g. https://glotzerlab.engin.umich.edu/photonics/index.html, https://www.nature.com/articles/s41467-021-22809-6, etc)
If you are interested in research in this direction, you can also have a look at other related publications, those results we will use intensively in the future: see e.g. https://www.nature.com/articles/s41467-022-31915-y and references therein. We always need a help from enthusiastic students, and have various possible projects to be involved in.
Fig.1 A scheme of ‘VAE+predictor’ ML model. VAE part of the model encodes images of photonic structures into latent space, and decodes them into the reconstructed structures.
Additional band predictor (or, in some versions, bandgap predictor) is trained to predict photonic banddiagram or bandgap, correspondingly, from the latent space. All three parts
(encoder, decoder, and banddiagram/bandgap predictor) are trained simultaneously.
Fig. 2 A scheme of latent space optimisation: shown are 6 structures taken from the dataset, which are then encoded into the latent space and move through the space according to a gradient descent algorithm with the objective to increase the bandgap. In the end of the process, the point in the latent space is decoded into a new, optimised structure.
Contact:
Dr. Alexander Itin alexander.itin(at)tuhh(dot)de