Abstract: The particle size distribution of suspended sediment is a key parameter in the study of water movement regulation and water conservancy construction. By combing the study of prior information such as Epstein-CarhartAllegra-Hawley (ECAH) model and ultrasonic attenuation experiment with the machine learning algorithms, the particle size of suspended sediment can be predicted. Features are extracted from the ultrasonic attenuation experiments and other related physical parameters, and labels are the particle size distributions determined by the sieving method. The training data sets and the verification data sets are made by the features and labels. The multi-output regression algorithm is constructed by combining the gradient boosting decision tree (GBDT) algorithm of single particle size prediction to predict the particle size distribution. The results show that the maximum relative errors of the single particle size of three samples are within ±10%, and the median diameter errors are 0.07%, -0.10% and -2.20% respectively. The predicted distributions are consistent with the results of the sieving method in the experimental range, which shows that the method has high feasibility and accuracy and can provide a new idea for particle size distribution measurement.