Informed Deep Learning for Electromagnetic Scattering Using Quasinormal Modes

Paper in proceeding, 2025

Neural networks have proven to be a powerful tool in the inverse design of meta-materials, metasurfaces and photonic crystals. However, a major limitation of neural networks is that they require very large amounts of data to train. We propose a network architecture for electromagnetic scattering that learns to represent scattering spectra in terms of resonances. We show that this approach significantly reduces the required amount of training data compared to standard neural networks. Our model is based on the highly general quasinormal-mode formalism, making it applicable to a wide range of devices.