Abstract: The corrosion resistance of zirconium alloy is of great relevance to the long-term operation of reactors. The surface of zirconium alloy undergoes oxidation in the complex environment of nuclear reactors, and their weight increases gradually. In order to investigate the corrosion weight gain behavior of zirconium alloy, the use of machine learning-based corrosion prediction methods for zirconium alloy provided new help. Firstly, long-term corrosion experiments of zirconium alloy under different conditions were conducted and the experimental data (corrosion temperature, pressure, media type, sample type and time) were recorded. The collected raw data were subjected to data preprocessing (missing value filling, deletion, standardization and normalization of the dataset). Then feature selection of the dataset was performed with the aim of performing correlation analysis and importance ranking of the features in order to take out the input features that have no effect on the output features and to reduce the dimensionality of the features to make the prediction model more accurate (avoiding overfitting and underfitting). Seven classical machine learning algorithms were used to establish the zirconium alloy corrosion weight gain model, and by comparing the differences in performance indicators, three models were selected for hyper-parameter optimization and cross-validation, and finally the optimized XGBoost-zirconium alloy corrosion weight gain prediction model was obtained. The results show that the model has high accuracy (R²=0.996) in the test set. To verify the stability and generalization ability of the above XGBoost-zirconium alloy corrosion weight gain prediction model, experimental data validation was carried out. A zirconium alloy sample heat-treated at 800 °C for 3 h was selected for corrosion experiments at a temperature of 360 °C, a pressure of 18.6 MPa and a corrosion medium of 100 ppb dissolved oxygen. Subsequently, the prediction was performed using the above model, and the R² in the validation dataset is 0.97, the RMSE is 6.76, and the average absolute percentage deviation of prediction is 5.3% approximately, which proves the prediction accuracy and generalization ability of the model. In order to improve the model interpretability of machine learning, SHAP explanation was used to illustrate how the input features affect the corrosion weight gain, where the corrosion time and temperature features are of high importance. From the analysis of the corrosion time and temperature dependence plot on each other, it can be seen that as the corrosion time increases, its contribution to the output feature (corrosion weight gain) is higher, i.e., the value of corrosion weight gain is higher. Key features are provided for further corrosion studies. The model developed in this paper can accurately predict the corrosion weight gain of zirconium alloy from a data-driven point of view, thus aiding and accelerating zirconium alloy corrosion studies.