Fiber Channel Modeling for Coherent Optical Fiber Communication Systems Using Koopman Neural Operator

Optical fiber channel modeling plays a vital role in the simulation, design, and performance assessment of optical fiber communication systems. Here, a new deep learning architecture, the Koopman neural operator (KNO) based network is investigated to learn the transfer function of a standard single fiber (SSMF) involving the joint effects of attenuation, chromic dispersion (CD), and nonlinearity. For effective channel modeling in long-haul coherent optical communication systems, we modify the loss function, specifically re-design the input vector for polarization division multiplexing (PDM) signals, and optimize the network parameters as well. The effectiveness of the model is evaluated in 28 GBaud PDM-16QAM/32QAM/64QAM coherent optical communications systems transmission over 1600 km, 800 km, and 400 km SSMF, respectively. Results show the normalized mean square error (MSE) between KNO-based model and split-step Fourier method (SSFM) is below 0.0035 over 1600 km transmissions at launch power range from -3 dBm to 6 dBm. Comparing the computational complexity of KNO with 0.1 km per step SSFM, the real multiplications (RMs) per symbol per polarization is only ∼0.9% of SSFM and the time consumption is ∼0.01 seconds over 1600 km transmission, versus ∼9.82 seconds for SSFM under the same hardware and software conditions. This scheme also shows robust generalization abilities in optical launch powers and modulation formats.