Inverse design of discrete Raman amplifiers using an invertible neural network for ultra-wideband optical...

The increasing demand for higher data transmission rates in optical networks has led to significant advancements in hollow core fibers (HCFs), particularly in multi-band transmission (MBT) systems. These systems, which use the previously untapped spectrum, require efficient amplification solutions to achieve stable long-distance transmission. Raman amplifiers, and especially discrete Raman amplifiers (DRAs), have emerged as promising candidates due to their broad bandwidth and tunable gain. However, optimizing the complex multi-pump configurations in DRAs remains a long-standing challenge. In this study, we propose an inverse design method based on an invertible neural network (INN) to efficiently determine DRA pump configurations for specific gain profiles. By combining a global search via the INN and local refinement through a fully connected neural network, our method achieves precise control over the gain flatness and pump parameters. Experimental validation over a 10 km spectral transmission system shows an 18 dB gain level with a gain flatness of 2 dB, confirming the effectiveness of the proposed inverse design method. In addition, the feasibility of the proposed design in practical applications was confirmed in the 25GBaud 64-QAM data transmission experiment based on HCF. This approach offers what we believe to be new opportunities for the optimisation of wideband optical networks.