This thesis presents machine learning models to predict the performance of volume visualization applications. The work focuses on two aspects of performance prediction related to volume visualization: the prediction of the execution time of an upcoming frame during runtime of an interactive volume visualization application and the prediction of the average performance of rendering volume data sets on arbitrary graphics cards. For dynamic frame time prediction, a volume ray caster with acceleration techniques is implemented, which allows user interactions. Data from the corresponding acceleration algorithms is used for the creation of a linear regression machine learning model, among other features. This model enables a real-time prediction of execution times needed for upcoming frames with a coefficient of determination between 0.67 and 0.96 for tested data sets. Predicting the average execution times on different GPUs is approached from two different directions. In the first one, a machine learning model is employed that allows the prediction of average execution times of an unevaluated data set. In addition to attributes describing the GPUs on which the performance is to be predicted, it only uses one feature specifying the volume, namely its file size. For tested data sets with high resolutions, predictions with a coefficient of determination between 0.56 and 0.83 could be made. In the second approach, a linear regression model is used which can predict the average execution time of a volume data set on an unevaluated system. Thereby, relative prediction errors between 5.33% and 22.22% on average could be achieved on different evaluated NVIDIA GPUs.