Dresden 2026 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MA: Fachverband Magnetismus
MA 34: Computational Magnetism I
MA 34.6: Vortrag
Mittwoch, 11. März 2026, 16:15–16:30, HSZ/0004
Material Parameter Determination with Deep Learning — •Jack Smith1,2, Dieter Suess1, and Florian Slanovc2 — 1University of Vienna, Vienna, Austria — 2Silicon Austria Labs, Villach, Austria
Measuring a set of material parameters from stray field measurements presents a challenging inverse problem. In this work, we introduce a data-driven approach for this task using a probabilistic neural network.
A standard regression approach can fail to capture the non-uniqueness of this task, leading to unphysical predictions. To mitigate this, we model the conditional probability distribution of the material parameters and measured stray field. We use a mixture density network (MDN)[1], which combines a neural network with a mixture density model, to model this conditional probability distributions. To capture spatial correlations, we employ a convolutional neural network feature extractor that projects the stray field map into a high-dimensional latent space before probabilistic decoding with the MDN.
We demonstrate that this method can extract ground-truth parameters with high accuracy, even for visually indistinguishable stray fields. It also provides a quantifiable uncertainty, identifying regimes where the stray field solution is non-unique and traditional inversion methods would fail.
[1] C. M. Bishop, Mixture density networks (1994).
Keywords: Inverse problems; Parameter Measurement; Deep Learning; Micromagnetics