Abstract: Leakage is a significant safety hazard for embankment dams, and accurate identification and localization of leakage inlets are crucial for reducing dam risk. Using simulated tracer distribution and transport data of leakage inlets, we trained a Transformer model to determine optimal parameter conditions and evaluate its predictive performance. The model's reliability was further verified through indoor experiments. The study found that: (1) when the number of iterations reaches 600, the relative error in the predicted maximum flow velocity is minimized, and the coordinates of the predicted maximum flow velocity are closest to those of the actual leakage inlet. Under this condition, when the data collection time is 50 seconds, the relative deviation in the predicted maximum flow velocity is also minimized, resulting in the best prediction. (2) With the optimal number of iterations and data collection time, the model achieves more than 95% prediction accuracy. The predicted values for leakage inlet size and flow rate are close to the actual values, with low relative errors in flow rate and location predictions. The relative error in location prediction is less than 5%. (3) Data from conductivity tests were converted into tracer concentrations and input into the model to predict flow velocity distribution. The model accurately locates the leakage inlet, with average relative errors in predicting flow velocity and inlet coordinates below 10%, confirming the model’s validity and accuracy in locating leakage inlets.These findings lay the theoretical foundation and provide technical support for the accurate identification of leakage inlets in embankment dams.