Deep Inverse Design of an Infrared Metasurface Diffuser

Machine learning (ML) algorithms have become invaluable tools for tackling design challenges associated with achieving unique scattering effects in artificial electromagnetic materials (AEMs). However, their effectiveness is reliant on substantial, well-constructed training datasets. Building such datasets using traditional methods becomes impractical for increasingly complex and large-scale geometric models. Achieving a specific diffuse scattering is one example and this often requires electrically large and diverse AEM arrays. Unfortunately, while numerical simulations offer high accuracy by utilizing fine meshing, their computational limitations render them incapable of handling such large structures and computing their scattering parameters efficiently. This work proposes a new approach to overcome these limitations by replacing conventional numerical simulations with a hybrid method that combines electromagnetic simulations with an analytical model, enabling the rapid and accurate generation of datasets for electrically large metamaterial arrays. Utilizing this approach, an optimized metasurface geometry for the mid-infrared range is successfully identified and tested that exhibits desirable diffuse scattering effects. This innovative method paves the way for significantly faster design and optimization of metamaterials, while also unlocking the potential for a new generation of large-scale, high-quality ML datasets for AEM problems.