The development of advanced nuclear reactor conceptions depends largely on the amount of available data to the designer. Non invasive ultrasonic techniques can contribute to the evaluation of gas-liquid two-phase regimes in the nuclear thermo-hydraulic circuits. A key-point for success of those techniques is the interpretation of the ultrasonic signal. In this work, a methodology based in artificial neural networks (ANN) is proposed to predict size distribution of bubbles in a bubbly flow. To accomplish that, an air feed system control was used to obtain specific bubbly flows in an experimental system utilizing a Plexiglas vertical bubbly column. Four different size distribution of bubbles were generated. The bubbles were photographed and measured. To evaluate the different size distribution of bubbles it was used the ultrasonic reflected echo on the opposite wall of the column. Then, an ANN has been developed for predicting size distribution of bubbles by using the frequency spectra of the ultrasonic signal as input. A trained artificial neural network using ultrasonic signal in the frequency domain can evaluate with a good precision the size distribution of bubbles generated in this system.