The use of levees is one of the most prevalent and effective strategies for flood protection. However

owing to the ageing of levees

inconsistent reinforcement efforts

and complex geological conditions

hazards such as piping frequently arise during flood seasons

which lead to significant and often irreparable damage. This study investigates backward erosion piping (BEP) in the foundations of double-structured levees via a back-propagation (BP) neural network optimized by a genetic algorithm (GA). The primary contributions of this study include: 1) the construction of a training dataset through numerical simulations of BEP in heterogeneous aquifers and validation of the dataset against laboratory sandbox piping tests to verify its reliability; 2) the extraction of head and permeability coefficient K data from Groups II

III

and IV in the BEP laboratory tests

augmentation of the dataset

and optimization of the GA–BP model to characterize test results in Group I

where the results demonstrate that the optimized model more accurately characterizes areas where the ≤1.0 cm/s; and 3) the use of the optimized GA-BP model to characterizes the development of a BEP channel. The results indicate that the model accurately captures the general trends. However

minor discrepancies remain in the characterized channel location and size compared with the actual conditions. In conclusion

this study offers an effective tool for characterizing BEP and demonstrates the potential of the GA–BP network model for practical applications in this field.