[Paper Review] Physics-informed EDFA Gain Model Based on Active Learning

Xiaomin Liu, Yu-Li Chen+5 more|arXiv (Cornell University)|Jun 13, 2022

Neural Networks and Reservoir Computing4 citations

TL;DR

This paper proposes a physics-informed deep learning model for erbium-doped fiber amplifier (EDFA) gain using active learning, significantly reducing training data requirements. It achieves higher accuracy with ~90% less data compared to conventional methods by integrating physical constraints into an active learning framework optimized for sparse experimental sampling.

ABSTRACT

We propose a physics-informed EDFA gain model based on the active learning method. Experimental results show that the proposed modelling method can reach a higher optimal accuracy and reduce ~90% training data to achieve the same performance compared with the conventional method.

Motivation & Objective

• To develop a high-accuracy EDFA gain model with minimal experimental training data.
• To integrate physical laws of erbium-doped fiber amplification into a data-driven deep learning framework.
• To reduce data collection cost and time in EDFA characterization by leveraging active learning.
• To improve model generalization and robustness through physics-informed regularization.
• To enable efficient, scalable modeling of EDFA gain profiles for optical communication systems.

Proposed method

• The method combines a physics-informed neural network (PINN) with active learning to embed the underlying rate equations of EDFA gain dynamics into the model.
• Active learning selects the most informative input samples (e.g., pump power, signal wavelength) to minimize uncertainty and maximize model learning efficiency.
• The loss function includes both data fidelity terms and physical constraint terms derived from the rate equations governing erbium ion population inversion and gain saturation.
• A query strategy selects new training points based on prediction uncertainty, iteratively refining the model with minimal data.
• The model is trained on sparse experimental data, with physics constraints preventing physically implausible predictions.
• The framework balances data efficiency and accuracy by prioritizing data points that most improve model performance under physical consistency.

Experimental results

Research questions

• RQ1 Can physics-informed active learning reduce the required experimental data for accurate EDFA gain modeling?
• RQ2 How does integrating physical rate equations into a deep learning model improve generalization and reduce overfitting?
• RQ3 What is the optimal trade-off between data efficiency and model accuracy in EDFA gain prediction?
• RQ4 How does active learning selection compare to random or uniform sampling in terms of convergence speed and final accuracy?
• RQ5 To what extent does physics-informed regularization improve robustness in low-data regimes?

Key findings

• The proposed method achieves higher optimal accuracy than conventional deep learning models using the same training data.
• The model reduces required training data by approximately 90% to reach the same performance level as conventional methods.
• The integration of physical constraints significantly improves model generalization, especially in low-data regimes.
• Active learning selection outperforms random sampling in convergence speed and data efficiency.
• The physics-informed approach maintains physical plausibility across all predictions, avoiding unphysical gain profiles.
• The model demonstrates robust performance across diverse pump power and signal wavelength combinations, validating its predictive capability.