LAUSR

The inverse-designed photonic device, with the characteristics of high performance and ultra-high compactness, is suitable for on-chip photonics applications. The gradient-based algorithms have high convergence efficiency. However, they depend on the continuous independent variable, so they cannot be directly applied to the pixel-based discrete search methods. In this paper, we propose a gradient-probability-driven discrete search (GPDS) algorithm for photonics inverse design. The algorithm establishes a connection between the gradient and the discrete value set by introducing the method of probability sampling. As an intrinsic discrete search algorithm in which the values of pixels are selected from a finite number of the discrete set, no additional discretization process is needed. Compared with the traditional brute-force search (BFS) method and traditional gradient method, the probability sampling process of our proposed GPDS algorithm can improve device performance efficiently and provide better stability to the initial states. We illustrate several component designs which are commonly used in the silicon photonics platform, and the results show that the algorithm can achieve high-performance structures within fewer iterations and has the ability of multi-objective optimization. With good flexibility and manufacturing-friendly geometry control, the algorithms are potential to be a powerful tool in solving multi-objective problems.