<p>Magnesium phosphate cement (MPC) is a new, environmentally friendly cementitious material with good mechanical properties, but its properties are easily affected by the proportion. A multi-objective optimization method was proposed to determine the optimal mix proportion of MPC paste using a backpropagation artificial neural network (BPNN) to predict performance indices (PIs) and a genetic algorithm (GA) to optimize the mix proportion. Firstly, the effects of the molar ratio of magnesium oxide to phosphates (M/P), the mass fraction of compound retarder, and the water-to-binder ratio on the compressive strength, the splitting tensile strength, and the setting time were investigated by the methods of orthogonal experiment design (OED) and analysis of variance (ANOVA). A BPNN was trained to reflect the relationship between the mix proportion of raw materials and the performance of MPC paste. The BPNN could be used to predict the PIs of MPC paste for a specific mix proportion of raw materials. Then, the GA combined with the prediction model was applied to find the optimal mix proportion of raw materials for the demand PIs over a larger range, thereby avoiding the errors caused by the scattered OED samples. At the same time, the method completed the calculation of the mix proportion of raw materials corresponding to the demand PIs. Throughout the process, the BPNN model was directly invoked in the GA, enabling the automatic calculation of PIs and optimization of the raw material composition, which benefited the subsequent application of MPC.</p>