In order to address the challenges of complex process and low precision in traditional device modeling, based on the double hidden layer conjugate gradient back propagation neural network (CG-BPNN) and the double hidden layer Levenberg-Marquardt back propagation neural network (LM-BPNN), two small signal models are proposed and analyzed for the gallium arsenide (GaAs) pseudomorphic high electron mobility transistor (pHEMT) here. At first, the scattering parameters (S-parameters) of GaAs pHEMT are divided into training set and test set randomly. Experimental results show that the CG-BPNN is better than another S-parameters when predicting ImS12 with mean square error (MSE) of 1.0449e-05, while LM-BPNN predicts ImS12 with MSE of 3.0954e-06. Meanwhile, the MSE of CG-BPNN is higher than LM-BPNN when predicting all the S-parameters. In addition, it shows a smaller fluctuation range for the error curve of LM-BPNN, which is more stable than the CG-BPNN. Therefore, the double hidden layer LM-BPNN is the better choice to model the small signal of GaAs pHEMT.