Kyra Klos1, Karin Everschor-Sitte2, Friederike Schmid1
1 Johannes Gutenberg University, Mainz
2 University of Duisburg-Essen, Duisburg
Topological defects and their dynamics are of high interest in a wide range of physics fields [1].
Due to the multiscale character of those defect structures and their complex interaction pattern resulting from the large size of associated deformation fields around each core, numerically simulating a large number of them in full microscopic detail is computationally highly challenging.
To provide insight into the connection between the macroscopic (particle) description of a model with topological defects and the underlying microscopic structure, we propose the use of neural networks.
With a spin-dynamic simulated microscopic model as training data, we use a conditional generative adversarial network system [2] with Wasserstein-loss [3] to generate reasonable spin-configurations from given defect configuration inputs. To generate realistic spin configurations, we take a two-step approach. First, the network learns how to generate a spin configuration with a desired defect distribution. Then, physical constraints are used to make the obtained spin configuration more realistic.
References
[1] Mermin N. D., Rev. Mod. Phys. 51, 591, (1979)
[2] Mirza M. ; Osindero S., arXiv:1411.1784v1, (2014)
[3] Arjovsky M. et al., ICML, PMLR 70, 214, (2017)
[4] Goodfellow I. et al., NeurIPS, (2014)