Integration of machine learning models for enhancing radioactive waste management of disused sealed radioactive sources

This research presents a novel machine-learning approach for optimizing the radioactive waste management of Disused Sealed Radioactive Sources (DSRS) through advanced predictive modelling. The study leverages comprehensive data from the Radioactive Waste Treatment Facility (IPLR), combining 1,339 rows of real operational data with 9,994 rows of synthetic data to develop robust prediction frameworks. Employing advanced preprocessing techniques such as SMOTE and ADASYN, we implemented five classification models (Decision Tree, k-Nearest Neighbors (kNN), CatBoost, Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM)) as part of the Reuse Identification Classification Model, which categorizes the likelihood of reusing DSRS. Subsequently, three regression models (Ridge Regression, Lasso Regression, and Random Forest) were applied in the Long-Term Utilization Regression Model to estimate long-term usability based on decay trends and activity levels. Our findings reveal that kNN outperforms other classifiers, achieving an AUC-ROC of 0.987, while Ridge Regression and Random Forest yield nearly perfect R-squared values, demonstrating superior long-term prediction accuracy. This study shows that machine learning has the possibility to improve DSRS management by accurately predicting reuse opportunities and estimating long-term requirements. A combination of real and synthetic data has produced models that aid in providing a more operational and data-driven radioactive waste management scheme. The results are significant for policymakers and other stakeholders in making informed decisions to enhance the sustainability and safety of radioactive waste handling.