Improved inverse design of polarization splitter with advanced Bayesian optimization

As many silicon nanophotonic devices are polarization-dependent, a polarization beam splitter that divides TE and TM modes is an essential component for photonic integrated circuits. Various structures have been proposed for polarization splitters, but it is still challenging to simultaneously achieve low insertion loss, high extinction ratio, compact size and simplicity of fabrication. In this paper, we combine new machine learning methods with the principle of multimode interference to propose a novel design for a polarization beam splitter. Our design has low insertion loss (-0.17dB/-0.42dB) and high extinction ratio (-23.1dB/-23.4dB) at the central wavelength of 1550nm for TE/TM modes, with compact size of 2×19μm2 and fabrication constraints strictly satisfied. Furthermore, our design is a standard 220nm-thick single-layer device for the silicon-on-insulator platform, without any auxiliary structures, making it easy for fabrication. Since our design cannot be optimized by the most commonly used methods, we adopt several specialized techniques to Bayesian optimization for inverse design. In this paper, we also share these skills which are simple but effective, possible to solve much more complicated design problems than others. [Display omitted] •We propose a novel polarization beam splitter, designed with machine learning.•Our design has low loss, high extinction ratio, compact size, and fab tolerance.•Our design is easy, reliable, and fabricable as a standard 220nm single-layer device.•We share the useful skills from recent progress of Bayesian optimization.