Björn Ziebeil and Vidar Gimbringer will present their master’s thesis work

While many industrial processes have been extensively studied using machine learning, powder metallurgy remains relatively underexplored. This thesis investigates the potential of machine learning models to predict the characteristics of metal powders after a reduction annealing process using time-series furnace data. A supervised learning pipeline was developed to predict six target variables: three particle size fractions, two chemical composition properties, and green density, including data preprocessing, model training, and performance evaluation. Multiple families of machine learning methods were implemented, including linear models, tree-based ensemble methods, and sequence models. Hyperparameters were systematically tuned, and model performance was evaluated using cross-validated error metrics. Feature importance was analysed using SHAP to identify the most influential process variables for each prediction task. 
The results show that predictive performance varies across target variables, with no single model consistently outperforming the others. Overall, LSTM, XGBoost, and Elastic Net achieved the strongest results, while all model families demonstrated applicability and shared key influential features. These findings highlight the promising potential for further integration of machine learning in powder metallurgy processes. 

Supervisors: Mikael Nilsson (LTH), Ola Litström (Höganäs AB), Fredrik Görtz (Höganäs AB)

Examiner: Niels Christian Overgaard (LTH)