Few-mode fiber design for multiple-input-multiple-output-less mode division multiplexing by machine learning

Neural networks (NNs), random forests (RF), and extreme gradient boosting (XGBoost) are applied separately to inversely design the ring-core few-mode fiber (FMF) with a desired weakly coupled performance. We demonstrate the procedure of inverse designing of FMF via machine learning (ML) algorithms and evaluate the prediction accuracy of the above ML algorithms. Compared with RF and XGBoost, the NN performs the highest prediction accuracy. For the NN, RF, and XGBoost, the correlation coefficients of the mode effective index difference are 0.99993, 0.99857, and 0.99937, respectively. Subsequently, by utilizing the method of permuting feature importance ranking, we obtain the high-correlation fiber structural features with the mode effective index difference. Moreover, we analyze the effect of the minimum index difference between two adjacent modes (Delta n(eff.min)) on the structural parameters and get consistent feature attribution via permuting feature importance ranking above. Finally, we design a weakly coupled ring-core fiber that supports four modes (HE11, TE01, HE21, TM01) based on the NN algorithm, which could be successfully applied in vector mode division multiplexing. (c) 2022 Optica Publishing Group

Automated summary

This study applies neural networks, random forests, and extreme gradient boosting to inversely design a ring-core few-mode fiber with desired performance. The prediction accuracy of these machine learning algorithms is evaluated, and the neural network demonstrates the highest accuracy. The study also identifies important fiber structural features related to the effective index difference and analyzes the impact of the minimum index difference on structural parameters.