Abstract: Soil bulk density is becoming more and more significant in modern agricultural production and research. However, it has not drawn enough attention in both practical production and scientific research because it is expensive and labor-some to obtain the exact value traditionally. Fortunately, the development of mathematical modeling techniques enables researchers to get soil bulk density indirectly using PTFs (pedotransfer functions) recently. In order to study the feasibility of PTFs’prediction for soil bulk density, the author applied BP neural network algorithm to establish PTFs. The study materials were 2 kinds of soil samples with different physical parameter, which were made in lab environment. The data used included the bulk density, cone index and volumetric water content of the examined samples, to train and test BP networks. They were measured with the help of a dual-sensor penetrometer. In this study, 6 PTFs based on BP neural network algorithm were established and evaluated by the root mean square error (RMSE) and the determination coefficient (R²) in the platform of MATLAB2008a. 2 types of soil samples (clay and silt-loam) with different physical parameters were made as study materials. According to the results, PTFs using BP neural network algorithm could effectively predict the soil bulk density with data from a dual-sensor penetrometer. The R² of clay PTF for bulk density reached 0.6973, and the R² of silt-loam PTF for bulk density reached 0.6868. It also supported that soil bulk density prediction using BP networks was significantly related with soil type, while had nothing to do with penetration depth.