Accelerating finite-difference frequency-domain simulations of inverse designed structures in nanophotonics using deep learning

Abstract

The inverse design of nanophotonic devices is becoming increasingly relevant for the development of complex photonic integrated circuits. Electromagnetic first-order simulations contribute to the overwhelming computational cost of the optimization routines in established inverse design algorithms, requiring more efficient methods for enabling improved and more complex design process flows. Here we present such a method to predict the electromagnetic field distribution for pixel-discrete planar inverse designed structures using deep learning. Our model is able to infer accurate predictions used to initialize a conventional finite-difference frequency-domain algorithm and thus lowers the average time required for simulating the electromagnetic response of nanophotonic device layouts by up to 53% in iterative design process flows. We demonstrate the applicability of our deep learning method for the inverse design of photonic integrated powersplitters and mode converters, and we highlight the possibility of exploiting previous learning results in subsequent design tasks of novel functionalities via fine-tuning reduced data sets, thus improving computational speed further.