Abstract Photonic crystal fiber (PCF) architectures have garnered significant interest due to their versatile applications across various fields. However, the complexity of PCF designs and the computational demands of full vectorial finite element method (FV-FEM) simulations pose challenges in fully realizing their potential. While prior research has explored artificial neural networks (ANNs) to accelerate simulation predictions, proposed approaches were often limited by sample size and generalizability. In this manuscript we introduce conditional generative adversarial networks (CTGAN) to augment real datasets, facilitating more effective ANN training. We evaluate CTGAN’s performance by comparing the quality of augmented data with real data and assess the predictive accuracy of ANNs trained on both augmented and real datasets. Our results demonstrate enhanced predictive accuracy, with higher R 2 values for optical property predictions from structural parameters using the augmented dataset-trained ANN. Furthermore, the mean squared error (MSE) during ANN training decreased significantly (from 0.0051 to 0.0011), requiring fewer convergence epochs of only 88 compared to 114 for the real dataset. The proposed approach enables faster optical property predictions while reducing the required dataset generation simulations by up to 27.9 %.
Read full abstract