VSR Seminar with Two Talks

1st talk: Stabilizing magnetic order in low dimensions: some DFT insights

Speaker:
Dr. Gustav Bihlmayer, Peter Grünberg Institute, PGI-1

Abstract:

To a large extent, storage and processing of information relies on the stability of magnetic entities representing different "states" required by the respective application. Two strategies that are applied to maintain this stability in the process of downsizing these magnetic structures are traditionally in the focus of the project JCIFF013: (i) the optimization of magnetic anisotropy and (ii) the formation of topologically stabilized magnetic structures. Both mechanisms depend on spin-orbit coupling and are investigated here using relativistic vector-spin density functional theory (DFT) calculations. In this talk I will focus on case (i) where we identify magnetic systems with large magnetic anisotropies.

Many lanthanides are characterized by an open 4f shell that leads to a large orbital moment. Only in lanthanides it is possible to measure magnetic hysteresis down to the size of a single adatom at a surface. We explored the physics of these open shell atoms on graphene, bridging the gap from DFT to the quantum-spin properties of the adatoms like quantum tunneling of magnetization [1]. Other examples are lanthanides intercalated between graphene and magnetic [2] or heavy substrates in different concentrations.

In a second example I will show the formation of a Ni kagome lattice on a superconducting substrate, Pb(111), that serves as template for magnetic adatoms on this surface. Our calculations guide scanning tunneling microscopy (STM) experiments with superconducting tips [3]. From the STM simulations, Fe monomers and trimers were identified. Such magnetic nanostructures are intensively studied as they can host Majorana zero modes at their edges or enable a superconducting diode effect. I will also discuss optimization and parallelization issues of the FLEUR code that made these simulations possible on JURECA-DC CPU and GPU.

[1] J. P. Carbone, J. Bouaziz, G. Bihlmayer, and S. Blügel, Phys. Rev. B 108, 174431 (2023)

[2] P. M. Sheverdyaeva, G. Bihlmayer, E. Cappelluti et al., Phys. Rev. Lett. 132, 266401 (2024)

[3] Y.-H. Lin, C.-J. Chen, N. Kumar, T.-Y. Yeh, T.-H. Lin, S. Blügel, G. Bihlmayer, and P.-J. Hsu, Nano Lett. 22, 8475 (2022)

2nd talk: Foundation Models for RNA Biology: Insights from NucleicBERT

Speaker:
Prof. Dr. Alexander Schug, Jülich Supercomputing Centre, JSC

Abstract:

RNA remains a frontier in molecular biology: while proteins benefited from breakthroughs such as AlphaFold, RNA structure and function prediction still suffers from scarce experimental training data and resulting limited computational accuracy. Our project addresses this gap with NucleicBERT, an alignment-free foundation model trained on 30 million RNA sequences. The model captures long-range dependencies, clusters RNA families, and highlights functional motifs directly from sequence, providing interpretable insights into RNA biology.

Training such a model required 2 million GPU hours of pretraining on HPC systems, while fine-tuning for structural or functional tasks is comparatively inexpensive. Explainable AI analyses show how NucleicBERT “rediscovers” RNA biology, bridging abundant sequence corpora with scarce annotations. Looking forward, we plan to extend the approach with multi-modal inputs and generative capabilities. By combining modern AI paradigms with large-scale computing, this project demonstrates how foundation models can open new avenues for RNA research, from basic biology to RNA-targeted therapeutic design.