Transformer-based eye opening factor prediction in upgraded 32-channel DWDM networks

Eye opening factor (EOF) captures the combined effect of noise, dispersion, and nonlinearity at the receiver decision threshold. This work predicts EOF in a 32-channel DWDM system from physical link parameters alone, excluding optical signal-to-noise ratio (OSNR) and Q-factor from the inputs. A reproducible closed-form model combining Gaussian-noise nonlinear interference, amplified spontaneous emission, direct-detection Q-factor, and dispersion- and PMD-induced eye closure generates 5,000 samples over 432 configurations (80–320 km, 8–32 channels, −5 to +10 dBm, 10–40 Gb/s, NRZ/RZ/DPSK). The link is upgraded with dispersion-compensating fiber (DCF), Raman amplification, variable optical attenuator (VOA) equalization, and idealized nonlinear compensation. Six regressors are compared. A Transformer with multi-head self-attention achieves the best held-out performance ( R 2 = 0.9928, RMSE = 0.0180). It is statistically tied with support vector regression under tenfold cross-validation and retains R 2 = 0.9930 when entire link configurations are withheld. Exact Shapley attribution identifies bit rate and residual chromatic dispersion as the dominant EOF determinants. A component ablation at 320 km shows that DCF is the enabling component, raising EOF from 0 to 0.745 at 10 Gb/s. At 40 Gb/s, only DPSK sustains BER < 10 −9 over 320 km, whereas NRZ and RZ reach this limit near 302 km and 274 km, respectively.