Berlin 2024 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MA: Fachverband Magnetismus

MA 6: Molecular Magnetism

MA 6.11: Talk

Monday, March 18, 2024, 12:30–12:45, EB 301

Deep Learning based Inverse Design of Ligand-Field Parameters of Single-Molecule Magnets — •Zayan Ahsan Ali, Julius Mutschler, and Oliver Waldmann — Physikalisches Institut, Universität Freiburg, D-79104 Freiburg, Germany

Single molecule magnets (SMMs) have attracted a rich volume of research in the recent decades due to their potential applications in magnetic memory and quantum computing. Lanthanide-based SMMs in particular demonstrate promising magnetic retention due to large inherent anisotropies. Their magnetic properties can be parameterized by ligand-field theories involving a set of 27 parameters. Experimental data such as magnetization and susceptibility curves, however, are typically featureless for these materials. Multiple distinct parameter sets can describe the data to equal accuracy, making the determination of model parameters a formidable inverse problem. In this work, the over-parameterization is tackled by implementing a deep learning architecture consisting of a Variational Autoencoder (VAE) in conjunction with an Invertible Neural Network (INN). The VAE-INN architecture determines hidden system parameters of the magnetic data and subsequently relates them to multiple valid model parameters from ligand-field theory. This approach is found to offer significant advantages over conventional fitting routines, such as Levenberg-Marquardt, in terms of generalization and convergence. The study investigates and presents both the merits and the effectiveness of the VAE-INN model in producing consistent sets of ligand-field parameters for novel experimental data.

Keywords: Machine Learning; Single Molecule Magnets; Inverse Problems; Neural Networks