The goal of this thesis is to find a way for choosing an optimal kernel for the data in multifidelity machine learning problem and then creating a model that will recreate the information contained in low-fidelity data and provide it forward. The model will be created by
using Gaussian process regression. Data that will be used is a wheel-soil interaction data, with the task of predicting the output force of a rover machine. The scarcity of high-fidelity data in a problem that leads to attempts to enhance models based on lower-fidelity data, which is easier to obtain. The ability to model lower-fidelity data with surrogate models is something that can be extremely helpful in achieving the best results possible. This approach can help advancing research in rovers movement, because it will eliminate part of the costs, by allowing for a better usage of multi-fidelity. The dataset has multiple overlapping features present, in order to address the problem of performance and combat memory dimensionality reduction techniques are implemented. We will try to use different methods to choose the best possible kernel for given data. Those will be based on the data and its specifics. In order to get results as good as possible at the later phases we will look into combinations of kernels. The final models that are benchmarked and compared according to their accuracy, the number of required data points, and possible further improvements.