Browsing by Subject "machine learning potential"

The rapidly increasing global energy demand has led to the necessity of finding sustainable alternatives for energy production. Fusion power is seen as a promising candidate for efficient and environmentally friendly energy production. One of the main challenges in the development of fusion power plants is finding suitable materials for the plasma-facing components in the fusion reactor. The plasma-facing components must endure extreme environments with high heat fluxes and exposure to highly energetic ions and neutral particles. So far the most promising materials for the plasma-facing components are tungsten (W) and tungsten-based alloys. A promising class of materials for the plasma-facing components is high-entropy alloys. Many high-entropy alloys have been shown to exhibit high resistance to radiation and other wanted properties for many industrial and high-energy applications. In materials research, both experimental and computational methods can be used to study the materials’ properties and characteristics. Computational methods can be either quantum mechanical calculations, that produce accurate results while being computationally extremely heavy, or more efficient atomistic simulations such as classical molecular dynamics simulations. In molecular dynamics simulations, interatomic potentials are used to describe the interactions between particles and are often analytical functions that can be fitted to the properties of the material. Instead of fixed functional forms, interatomic potentials based on machine learning methods have also been developed. One such framework is the Gaussian approximation potential, which uses Gaussian process regression to estimate the energies of the simulation system. In this thesis, the current state of fusion reactor development and the research of high-entropy alloys is presented and an overview of the interatomic potentials is given. Gaussian approximation potentials for WMoTa concentrated alloys are developed using different number of sparse training points. A detailed description of the training database is given and the potentials are validated. The developed potentials are shown to give physically reasonable results in terms of certain bulk and surface properties and could be used in atomistic simulations.