Westinghouse Electric Sweden is an established company specializing in nuclear fuelproduction, where maintaining high and consistent fuel quality is essential for reactor safety andoperational reliability. The production process is complex, involving multiple interdependentstages, including conversion and pellet manufacturing, where uranium hexafluoride is convertedinto uranium dioxide powder and subsequently pelletized into small cylindrical fuel units.

This thesis investigates the relationships between conversion process parameters and powdercharacteristics, and between pelletization process parameters, powder characteristics, andpellet properties using regression analysis. Additionally, the objective is to develop machinelearning models that predict the optimal: amount of recycled uranium dioxide (addback);pressed pellet density; press process parameters, and powder characteristics.

Historical production data was collected, preprocessed and analyzed using correlation analysisprior to subsequent modelling using supervised machine learning. Linear regression was usedas a baseline, while neural networks, gradient tree boosting regression, and gaussian processregression were evaluated with a target prediction accuracy of 𝑅² = 0.85.

The results show that individual parameters exhibited weak correlations due to the complexity ofthe production process, while multivariate models captured some underlying relationships. Thebest performing model for addback prediction was XGBoost, exceeding the target accuracy.However, green density and pressing parameters proved more difficult to predict, with allmodels showing weaker predictive accuracy. Low to moderate predictive performance wasachieved for the conversion-based models for powder characteristics, although the desired levelof accuracy was not reached.

Overall, the study demonstrates that machine learning can capture important processrelationships, though predictive performance is highly dependent on data quality and processcomplexity.