A back propagation (BP) algorithm of the neural network was established so as to use Molecular Interaction Volume Model (MIVM) to predict the components activity in alloy solution accurately and conveniently. Its input layer is the experimental activity coefficient of components in alloy solution, output layers are molecular pair energy interaction parameters and hidden layer is set to one layer. The structural parameters in BP neural network model were optimized via genetic algorithms, where the deviation between the experimental and theoretical values of the infinite dilution activity coefficient as fitness function. The optimization was carried out with minimum deviation so as to ensure the validity of BP neural network. Finally, the BP neural network model and algorithm were validated by taking the binary alloy solutions of Pb-Bi, Sn-Bi, Sn-Pb and Fe-Cu as examples. The results show that the average relative errors between the predicted and experimental values of components activity in the alloys are less than 4% and the absolute deviation less than 0.78, which can meet the requirements of engineering calculation.