Inverse design of photonic crystal fibers for dispersion engineering using neural networks

This work presents a novel approach for the inverse design of photonic crystal fibers, by implementing neural networks to engineer dispersion profiles. We investigate two neural network architectures — inverse neural networks and tandem neural networks — across three configurations: a direct inverse neural network, a tandem neural network, and an optimized inverse neural network. Our results show that the optimized inverse neural network achieves superior performance without the need for a complex tandem architecture. By using five strategically selected data points from the dispersion spectrum, the design process is streamlined, enabling accurate predictions with minimal input data. Neural network predictions are validated through finite-difference frequency-domain simulations and further confirmed by experimental chromatic dispersion measurements. Additionally, we evaluate the tolerance error on the dispersion curve due to defects introduced during the stack-and-draw process and demonstrate the versatility of the neural networks in modeling various physical phenomena with fast and straightforward designs. While focused on photonic crystal fibers, this methodology can be extended to other photonic structures, such as integrated waveguides optimized for specific optical properties. Our findings highlight the potential of neural networks for efficient, precise inverse design in diverse photonic applications.