Large-scale three-dimensional (3D) physical modeling is an important method to study landslide-induced impulse waves. In such models, the test randomness is often quite high, which necessitates systematic exploration of the randomness and error. However, only a few relevant studies have been conducted yet. To this end, this study aims to investigate the randomness and error of large-scale 3D landslide-induced impulse wave experiments and provide solutions to the different sources of error. Based on six repeatability experiments with the large-scale 3D physical model of the Wangjiashan landslide-induced impulse wave in the Baihetan reservoir of the Jinsha River, China, the errors of typical physical parameters are classified into systematic errors, which originate from instrumental factors, experimental design, observer bias, environmental factors, and random errors originating from communication and observation. The allowable error rate of landslide motion in the repeatability experiment is found to be 5%, but the dynamic chain transmission of landslide-induced impulse waves leads to the transmission and accumulation of errors, which causes a gradual increase in the errors of landslide motion, primary wave, propagating wave, and run-up process; and the coefficient of variation increases from approximately 3.8% to 25.0%. To reduce the experimental data error, a low-pass filtering model for removing high-frequency noise and a moving window smoothing model for image frame rate mutation are established, which can decrease the coefficient of variation by nearly 1.3%-4.0%. The corrected particle dynamic map exhibits a continuous and smooth flow field, which basically eliminates the velocity field mutation and discontinuity caused by communication data packet loss. Overall, this study can provide theoretical basis and technical support for large-scale 3D landslide-induced impulse wave experiments.
