Accelerated inverse design of passive Si Photonics

We present an accurate differentiable parametrization scheme for topological optimization and inverse design of passive components of integrated Si photonics. We show that the most of operations in the transformation chain from control variables to effective anisotropic dielectric tensor discretized on the Yee grid are expressible in terms of generalized convolutions that can be computed efficiently on GPU accelerators. Combined with the recent open-source GPU FDTD code and custom gradient-descent optimizer with dynamic control of final design manufacturability, it results in fast inverse design tool capable of producing reliable Si photonic component layouts in minutes/hours instead of days. The final designs, although not globally optimal, are generally physically meaningful since they reflect how fields themselves tend to propagate through the device in the most efficient way.