Contacts: Associate professor Kaspar Kirstein Nielsen (kaki@dtu.dk) or associate professor Rasmus Bjørk (rabj@dtu.dk)
Challenge: A lot of hype currently exists about the usage of artificial intelligence, neural networks and deep learning. At the same time, in the field of permanent magnets there is room for improving algorithms for optimizing permanent magnet structures (in terms of resulting magnetic field, homogeneity, field gradient etc.). The challenge in this project is to apply deep learning techniques to do these optimizations and perhaps in the end to come up with an unexpected solution.
Idea: The idea is to train a neural network (NN) using an existing in-house magnet numerical model. The model essentially consists of “lego-bricks” of magnet material that can be placed in any position and orientation. The output of the model can then be the magnetic field in a specific region, the homogeneity of the field, the force between to magnetized bodies etc. Training the NN with a large number of random runs of the model (within the bounds of a well-defined problem) should then yield an NN that can recognize such configurations. Further along, the NN should be inverted, i.e. the user should be able to ask the question: “What does a magnet configuration producing a field of X tesla look like?” and the NN should then come up with a design suggestion.
Students’ task: You will be tasked with setting up a training environment for the NN, train it and evaluate how well it can recognize new configurations (and compare this to the model predictions). Inversion of the NN is most likely a future task, but certainly something, that would be relevant to work on.