The CdS/CdTe heterojunction plays an important role in determining the energy conversion efficiency of CdTe solar cells.However,the interface structure remains unknown,due to the large lattice mismatch between CdS and CdTe,posing great challenges to achieving an understanding of its interfacial effects.By combining a neural-network-based machine-learning method and the stochastic surface walking-based global optimization method,we first train a neural network potential for CdSTe systems with demonstrated robustness and reliability.Based on the above potential,we then use simulated annealing to obtain the optimal structure of the CdS/CdTe interface.We find that the most stable structure has the features of both bulks and disorders.Using the obtained structure,we directly calculate the band offset between CdS and CdTe by aligning the core levels in the heterostructure with those in the bulks,using one-shot first-principles calculations.Our calculated band offset is 0.55 eV,in comparison with 0.70 eV,obtained using other indirect methods.The obtained interface structure should prove useful for further study of the properties of CdTe/CdS heterostructures.Our work also presents an example which is applicable to other complex interfaces.