Distribution Ratio Prediction of Major Components in 30%TBP/kerosene-HNO₃ System Based on Machine Learning

Spent fuel reprocessing is an important nuclear energy process which aimed at recovering resources and managing radioactive materials to control potential hazards. In this field, Purex technology is widely used for its high efficiency, scalability, and wide applicability. Purex technology, a liquid-liquid extraction technique to separate and purify uranium and plutonium from nuclear fuel, plays a key role in spent fuel reprocessing, enabling reprocessing and recycling of nuclear fuel, reducing the release of radioactive nuclear waste, and improving the efficiency of nuclear energy resources. Meanwhile, as an emerging technology, machine learning has attached wide attention and has been applied in the field of Purex, such as the selection of ligands and ionic liquids, the prediction of ligand properties, and so on. In this paper, machine learning is combined with distribution ratio prediction, which is defined as the distribution ratio of ionic liquids in different phases, which can reflect the extraction rate of ions, and plays an important role in Purex computer simulation, so the distribution ratio prediction model can help researchers to choose the optimal experimental conditions, optimize the process, and reduce the experimental cost and time. Since the traditional mathematical model of uranium distribution ratio leads to at least 15% prediction error, in this paper, three classical machine learning models (namely, random forest, support vector regression and K-nearest neighbor) were constructed to predict the distribution ratios of uranium, plutonium, and HNO₃ in the 30%TBP/kerosene-HNO₃ system. These models were trained based on different datasets, and their hyper-parameters were optimized using algorithms such as grid search, Bayesian optimization, and K-fold cross-validation. The results show that random forest achieves the best results in distribution ratio prediction. The average absolute relative error (AARE) of uranium distribution ratio prediction reaches 7.73%, which is about 7% higher than that of the traditional model. In addition, plutonium and HNO₃ distribution ratios are also predicted to verify the generalizability of the machine learning model, and the highest of 11.6% and 13.7% are achieved. The machine learning model prediction results show that the machine learning method proposed in this paper achieves better performance than the traditional distribution ratio mathematical model, effectively improves the accuracy of uranium distribution ratio prediction, and performs well in plutonium and HNO₃ distribution ratio prediction.