This paper presents an estimate of surface layer optical turbulence in Northwest China using an artificial neural network. We optimize the configuration of the multilayer perceptron (MLP), including 10 features in the input layer and 40 neurons in the hidden layer. The performance of the constructed MLP is investigated. The results show that when the training set and testing set are from the same site, the mean relative error of the model is 1.34%. The goodness of fit between measured and estimated refractive index structure constants is 0.94. We propose that when the training set and testing set come from different sites, the generalization ability of the MLP should be enhanced.