Bruno Oliveira Cattelan participated at the AI Day 2023 with a poster presentation on the preliminary result of their machine learning work applied to material research.
Abstract:
Material research plays an important role in fusion. From deciding which material can withstand the high temperatures divertor components face to the best wall material to minimise damage and dust in case of irradiation effects. We focus on vacancy defect binding energies to screw dislocation. Traditional techniques such as density functional theory can for some configurations give the answer. However, in order to study the binding energy when vacancies are present, a combinatorial amount of cases need to be analysed, which quickly becomes infeasible.
To combat this issue, we present a neural network solution. From a subset of cases we can train a model, which in turn can predict the energy in a fraction of the time. We make use of state of the art techniques such as Dropout and BatchNormalization to ensure the best possible performance of our model. However, due to the nature of the problem we have to deal with a large amount of uncertainty. This is addressed by using uncertainty quantification techniques such as deep ensembles and mixture density networks. We show that our solution has a mean validation error around 5.7%, with whole datasets predicted in a matter of minutes.
Full list of authors:
Oliveira Cattelan, Bruno (University of Helsinki); Lindblad, Victor (University of Helsinki); Granberg, Fredric (University of Helsinki)