제목 : Accelerating photonic inverse design through adjoint optimization and deep learning
초록 : The rapid evolution of inverse design methodologies, driven by large-scale optimization across all geometrical degrees of freedom, is transforming the field of photonic design. Despite this progress, these methods rely heavily on full-wave Maxwell solutions to compute gradients for desired figures of merit, resulting in substantial computational burdens on traditional platforms. This talk introduces cutting-edge strategies to mitigate these challenges, such as neural network-based electromagnetic solvers and adjoint-based data augmentation techniques. While neural network-based solvers (surrogate solvers) excel at inference under fixed conditions, they often require retraining for varying scenarios. To address this, I will present the Wave Interpolation Neural Operator, a novel surrogate solver that enables an inference for untrained simulation conditions. Additionally, I will explore adjoint-based data augmentation, which predicts figure-of-merit changes due to structural modifications using only two full-wave simulations. By leveraging adjoint gradients, we can generate and label thousands of new data points without additional computations. Finally, I will highlight recent advancements in deep-learning-assisted metalens imaging.
