基于机器学习的锆合金在360℃/18.6 MPa溶氧水中腐蚀预测方法研究

Research on Machine Learning-based Corrosion Prediction Method for Zirconium Alloy in Dissolved Oxygen Water at 360℃/18.6 MPa

锆合金在核反应堆复杂环境中会发生氧化,为进行高温锆合金腐蚀增重研究,以基于机器学习的预测方法为其提供新的辅助手段。数据集来源于实验记录数据,使用7种机器学习算法进行锆合金腐蚀增重模型的建立,通过对比性能指标的差异,进行进一步的超参数优化和交叉验证,最终得到优化的XGBoost-锆合金腐蚀增重预测模型。结果表明,经过验证集进行验证得出预测值的平均绝对百分比误差约为5.3%。通过SHAP解释法进行模型的解释性分析,得出重要性较高的特征为腐蚀时间和温度,为进一步的腐蚀研究提供了关键特征。最终所建立的模型可以从数据驱动的角度准确预测锆合金腐蚀增重情况,辅助和加快锆合金腐蚀研究。

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(R^(2)=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℃ for 3 h was selected for corrosion experiments at a temperature of 360℃,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.