Photonic digital gates are the next generation of all optical digital devices. Exclusive OR gate (XOR) is one of the important and applicable components in the next generation of all-optical networks. Generally, the optical digital gates are simulated by complex and time- consuming numerical methods like FDTD. Prediction of the accurate output (the state of 0 or 1) is a key parameter in the digital gates. Photonic devices modeling by artificial neural networks (ANNs) will be introduced as a flexible, suitable and precise modeling alternative approach instead of numerical simulations. In this paper, performance of a 3-input all-optical exclusive OR gate (XOR) has been modeled using artificial neural networks. We discuss in detail that trained ANNs can be used as a fast modeling method for optical digital gates, with high accuracy. Here, we proposed to develop a modeling system for alloptical 3-input XOR gates based on multilayer perceptron (MLP) and radial basis function (RBF) neural networks. Therefore, the proposed neural networks are trained by data arising from simulations by numerical methods. The dataset is the power lasers of all possible logical conditions used to train neural networks. The full data set is split in 90/10 for the train and test data. To understand the ANN methods' performance, the figures of the predicted results of all optical XOR gate are plotted and compared. Also, the error indices like mean square error (MSE) and Relative Square Error (RSE) are calculated for the test data to evaluate the performance of neural network models on the prediction of all optical XOR gate output. The correlation between the modeled data by neural networks and simulated data by numerical methods are established by Correlation Coefficient (R-2) parameter too. According to the comparison of both neural networks algorithms, good results are reached which lead to an effective technique. The effect of training parameters of ANN' models, like number of hidden layers, number neurons in the hidden layers, number of epochs, learning rate value, spread of gaussian functions on the prediction results and errors are compared and analyzed. The aim is to set the optimum parameters to prevent from complexity of the neutral network. The best results for RBF NN is the value of 1 for spread of Gaussian function and 90 hidden neurons that lead to MSE, RSE and R-2 values of 4.0837 x 10(-4), 0.0114, and 0.9888 respectively. The optimum structure of MLP NN is 2 hidden layers with 12 and 8 neurons (5 12 8 1) by training with 65 epochs. The activation function for hidden and output layers are chosen logsigmoid, logsigmoid and purelin respectively. The calculated MSE, RSE, and R-2 for the best MLP NN structure has been 7.5 x 10(-10) 0.000133, and 0.9999, respectively, which confirms the high accuracy of the mentioned neural network model. Even though the two mentioned neural networks models can be used to model all-optical 3-input XOR gates appropriately, results show that MLP NN using the most relevant features achieved the best results and better estimates. Finally, the implementation of the optical logic gates with neural network was illustrated for the future optical integrated circuits.
