Always-Feasible Photonic Inverse Design with a Differentiable Conditional Design Generator

Inverse design has become an effective automation tool for generating high-performance, fabrication-feasible photonic integrated devices, enabling the manipulation of light in multiple dimensions. However, due to the incorporation of fabrication constraints such as minimal feature size or spacing into the continuous optimization process, conversion from an optimal yet infeasible design topology obtained from computational algorithms to a physically reliable one has presented a challenge, potentially compromising its optimality or leading to increased optimization iterations. In this work, we propose the use of a bilevel optimization algorithm to address the fabrication-constrained inverse design. The inner-level optimization serves as a differentiable feasible design generator, while the control variable of the design generator is optimized in the outer-level problem. This approach enables the precise acquisition of the gradient of a desired figure of merit, thereby eliminating the need for gradient estimation with robust convergence properties. Governed by the always-feasible framework, all of the intermediate devices on the optimization trajectory can adhere to the fabrication requirements. We validate the effectiveness of our method through optimization tasks for various photonic integrated components using both 2D and 3D simulations. The optimized designs are also fabricated and characterized in the experiment. Our results from simulation and experiment highlight the benefits of our new method in designing high-performance and reliable integrated photonic devices that satisfy fabrication limitations.