Deep Learning Accelerated FDFD Simulations in Context of Inverse-Design Algorithms

Schulte, Lukas;Butz, Marco; Schuck, Carsten

Zusammenfassung

Deep learning (DL) methods have shown tremendous success in various disciplines related to the design of photonic integrated circuit components. Similarly, DL may be used to accelerate electromagnetic first-order simulations. To numerically access the photonic properties of metamaterials requires efficient electromagnetic simulation algorithms. In order to simulate the propagation of light through nanophotonic devices, various numerical methods, such as the finite-difference frequency-domain method (FDFD), have been developed. Requiring to solve large sparse linear systems iteratively, these methods come at the cost of being computationally expensive processes. Here, we show how DL can be employed to decrease the computational effort of consecutive FDFD simulations, for example, encountered in inverse-design algorithms. Leveraging the U-Net architecture our method is capable of predicting the electromagnetic response of a nanophotonic device. We use this prediction as a starting point for iterative refinement using FDFD simulation, thus decreasing the required iteration and computation time drastically. Hereby, we minimize the overall time required by inverse-design algorithms to reach convergence and thus enable more efficient and compact device layouts.