For high symbol rate fiber optic networks, the estimation and monitoring of time varying link performance parameters are critical for delivering optimal network performance. In this paper, a method is presented for estimating the nonlinear signal-to-noise ratio ${\text{(SNR}}_{{\text{nl}}})$ using an artificial neural network (ANN). The ANN is trained with the fiber nonlinearity induced amplitude noise covariance and phase noise correlation extracted from received symbols. The data used for training are simulation results for a 34.5 Gbaud dual polarization 16-ary quadrature amplitude modulation signal transmitted over a wide range of link configurations with varying fiber types and number of wavelength division multiplexed channels. Using 734 input-output sets, high accuracy is demonstrated for training, testing, and validation for simulation data with a maximum normalized root-mean-square error of 0.37% for ${\text{SNR}}_{{\text{nl}}}$ . Validation using experimental data exhibits less than 0.25 dB deviation from the true ${\text{SNR}}_{{\text{nl}}}$ for estimates obtained with varying fiber length and launch power.