In this paper, the performance of particle swarm optimization–radial basis function (PSO–RBF) neural network was examined to predict the shear strength of granular material. Direct shear tests were conducted on glass beads (400–600 mm in size) with 2% of liquid content at different powder contents (1, 2, 3, 4, and 5%) as the binder. Hydrated lime (HL), quick lime (QL), calcite, and sodium lignosulfonate (SL) powders were used to compare their shear strength characteristics. The actual loading stress was controlled at five levels (6.9, 10.3, 13.8, 17.3, and 20.7 kPa), and five further levels (24.2, 27.7, 31.2, 34.6, and 38.1 kPa) were used for prediction. Results showed that at 2% of water, the shear strength regularly increased at all amounts of HL, QL, and calcite powder. With the increased SL powder, the shear strength decreased. The actual and predicted maximum shear stress values for different levels of normal stress were fitted using the widely accepted Mohr–Coulomb criterion to determine the internal friction angle and cohesion, and R 2 was improved from 0.98 to 0.99. The PSO–RBF neural network model was a flexible and accurate method for the prediction of the shear strength of granular material.