Coronagraph design uses advanced optimization methods, specialized for each coronagraph, to solve the largescale and often non-convex optimization problems. Particularly this non-convexity makes reliably finding highthroughput solutions difficult. The optimization problems are therefore usually separated in independent stages, or individual masks are parameterized with a small number of hyperparameters, to make the problems tractable. We instead aim to solve the original full optimization problem directly. We use the Augmented Lagrangian Method to reduce the full constrained optimization problem into a series of unconstrained sub-problems. We then solve each sub-problem using a gradient-based non-linear optimization method with efficient analytical gradients obtained with algorithmic differentiation. We employ periodic relaxations to deal with the inherent non-convexity in the sub-problems. Computations are performed on GPUs for additional performance. We show examples of apodizing phase plate coronagraphs, apodized-pupil Lyot coronagraphs, apodized vortex coronagraphs, and phase-induced amplitude-apodization coronagraphs optimized with our generalized algorithm, confirming previously known state-of-the-art solutions for each. This work paves the way for exploration of more exotic optical layouts of future coronagraphs.
|