In order to quickly and accurately obtain magnesium oxychloride cement concrete （MOCC） proportions with excellent water resistance， a particle swarm optimization back propagation neural network （PSO-BPNN） model with a topology of 4-10-2 was designed. The input layer parameters of the above model were n（MgO）/n（MgCl₂）， fly ash content， phosphoric acid content， and phosphate fertilizer content. The output layer parameters were MOCC compressive strength and softening coefficient. The model establishment data set contained 144 groups， including 100 groups of training set data， 22 groups of validation set data， and 22 groups of test set data. The results show that the mean value of each evaluation parameter in the prediction of compressive strength using the PSO-BPNN model are coefficient of determination R²=0.99， mean absolute error S_MAE=0.52， mean absolute percentage error S_MAPE=1.11， and root mean square error S_RMSE=0.73. The mean value of each evaluation parameters in the prediction of softening coefficient are R²=0.99， S_MAE=0.44， S_MAPE =1.29， and S_RMSE=0.62. This indicates that compared to the BP neural network （BPNN） model， the PSO-BPNN model has a strong ability to predict dual parameters and can be used for both forward design and reverse guidance of MOCC mix proportions.