Neural Network (NN) is used to improve an Explicit Algebraic Reynolds Stress Model (EARSM). The turbulent kinetic energy and its dissipation are predicted using the standard \(k-\omega \) model. The NN model is trained in channel flow of \(Re_\tau = 10{,}000\) . The NN model is stored to disk and subsequently loaded into the CFD code. The NN model is called every iteration to compute the \(\beta \) coefficients in the EARSM, i.e. the CFD solver and the NN model are fully coupled. The Reynolds stresses are used in the momentum equations and the production term in the k and \(\omega \) equations. It is found that when training the NN model, the target data cannot only be taken from DNS. The reason is that the stress-strain relation and the turbulent kinetic energy of the DNS data are different from those of the \(k-\omega \) model. Hence, the target data are taken both from DNS and a \(k-\omega \) simulation. The new EARSM-NN model is used for predicting channel flow at \(Re_\tau = 2000, 5200\) and 10, 000 and flat-plate boundary layer at \(2500 \le Re_\theta \le 8000\) . The EARSM-NN model gives much better results than the standard EARSM.