In the metallurgic industry, aggregates are used to melt and treat different types of metal alloys, such as steel. In order to keep the vessel containing the molten steel from melting itself, the inner surface is coated with special refractory material. Several factors, such as the concrete composition of the steel being produced, numerous chemical additives and other process parameters expose the refractory lining to severe levels of wear and tear.

Eventually, this leads to vessels failing, which is not only dangerous but also expensive due to halt in production. Therefore, it is essential to have a good appraisal of the current level of wear of a vessel, in accordance to past treatment processes, so called heats. Since the precise factors to variations in wear are unknown, machine learning (ML) is applied to uncover latent patterns in the wear and tear process to make predictions. This thesis aims to use state-of-the-art ML algorithms, in particular neural networks to model the wear of refractory material used as inner coating in metallurgic vessels. A special challenge arises from the fact that measurement of the remaining thickness of refractory bricks is infeasible to be carried out after each heat and thus are only possible after the lifetime of a vessel. Different ML architectures like convolutional neural networks (CNNs) and long short-term memories (LSTMs) are discussed and evaluated. Here, two different sources of datasets were used. The first one was provided by Voestalpine and is referred to as Linz datasets. It contains wear measurement data from three different regions of a Ruhrstahl-Heraeus-degasser (RH degasser). The second stems from an undisclosed source and is referenced as Ladle. It contains two different datasets of wear measurements. An extensive hyperparameter tuning process leads to optimal architectures for each of the two ML approaches. It was shown that cosine restarts learning rate (LR) scheduling and dropout were substantial contributions to enhance wear prediction performance of the models. Several other optimization techniques were examined, such as providing the models with estimations of the wear after each heat, which lead to significant improvements. The outcome of the model evaluations led to promising results in the task of predicting refractory wear with deep learning. Both, LSTM and CNN models showed decent performance on all datasets, with the LSTM model standing out for being smaller and less complex.