Gas holdups and liquid circulation velocities in two external loop circulating bubble columns of the open channel gas separators using air-water and air-glycerol systems were extensively reported by Al-Masry (1999, 2004). The effects of changing the volume of the liquid in the gas-liquid separators on, the columns hydrodynamics were analysed numerically using neural network with four inputs and three outputs. The inputs were superficial gas velocity U-GR, volume ratio T-VR, liquid Viscosity mu(L) and scale-up factor A(D)/A(R), while the outputs were liquid circulation velocity U-LR, riser gas holdup epsilon(GR) and downcomer gas holdup CGD. The network was trained on 60% of the data, and then used to predict 40% of the data that have never been seen by the network. The training was successfully accomplished and results obtained with average normalized square error < 0.01. Comparison of the neural network predictions of the hydrodynamics variables with predictions of Al-Masry (2004) gave much better improvement. The results show that neural networks, if properly designed, are very powerful predicting mathematical tools that can accurately approximate nonlinear input-output mappings.