Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MA: Fachverband Magnetismus

MA 41: Poster III

MA 41.27: Poster

Donnerstag, 20. März 2025, 15:00–17:30, P3

Exploring Data Representation Techniques in Deep Learning Models for Determining Ligand Field Parameters of Single-Molecule Magnets — •Preeti Tewatia, Zayan Ahsan Ali, Julius Mutschler, and Oliver Waldmann — Physikalisches Institut, Universitat Freiburg, D-79104 Freiburg, Germany

Single-Molecule Magnets (SMMs) present an exciting frontier in molecular electronics and quantum computing. According to ligand field theory, the single-ion magnetic anisotropies have in general to be characterized by 27 ligand field parameters. However, typical experimental data such as magnetic susceptibility as function of temperature measured on powder samples is pretty featureless, leading to an inverse problem where multiple parameter sets can equally describe the data. To address this challenge, a deep learning approach based on a Variational Autoencoder and an Invertible Neural Network hybrid architecture was employed. The model has been demonstrated before to be capable of handling the above inverse problem. This work focuses on improving the results of the model using data representation and augmentation techniques. For instance, augmenting the input data with simulated susceptibility curves which include experimental errors were found to enhance the robustness of the model with respect to these errors. This approach leads to better performance than conventional fitting techniques.

Keywords: Machine Learning; Deep Learning; Ligand Field; Data Representation