Towards Autonomous Optical Fibre Networks: High-Precision EDFA Gain and Spectral Response Prediction via Hybrid CNN-LSTM Deep Learning

This paper introduces a hybrid CNN-LSTM architecture for automatic advantage prediction and optimization in erbium-doped fibre amplifiers (EDFAs), addressing a crucial role in maintaining signal strength over long-distance optical networks; however, existing modelling techniques face significant challenges in balancing accuracy and computational efficiency. The proposed version uniquely integrates convolutional neural networks (CNNs) for spatial-spectral function extraction and long short-term memory (LSTM) networks for temporal dynamics modelling, allowing simultaneous prediction of benefit profiles, 3 dB compression factors, and full-width half-maximum (FWHM) bandwidths from 5 input parameters: pump electricity, signal energy, fibre length, erbium concentration, and wavelength. When validated against OptiSystem simulations across 10 fibre lengths (3–30 m), the framework achieves unheard accuracy (R² >0.999, MSE=0.0032) while decreasing the computational time from hours to 6.1 milliseconds in line with the prediction—a 600,000× speed improvement. Benchmarking in the direction of seven contemporary strategies demonstrates 66–88% stepped forward average performance in essential metrics: 4× decrease in three dB point mistakes (0.18 dBm vs. 0.92 dBm in CNNs), three.4× better FWHM precision (0.36 nm vs. 1.21 nm) for 1 m--20 m as a fibre length, and real-time functionality with 1.28 million parameters. These upgrades permit self-reliant EDF optimization in dynamic optical networks, which remedy the spatiotemporal doped cloth that affects the modern-day process.