Harnessing LUMI: Polish researchers unveil atomic-level insights into catalytic materials

Understanding the electronic structure of complex catalytic materials requires simulations that go far beyond the capabilities of conventional computing. In materials like cobalt spinel-ceria (Co₃O₄|CeO₂) heterojunctions, their performance in chemical reactions depends on how atoms are arranged and how electrons move between different parts of the material.

To understand these processes in detail, scientists use quantum-mechanical simulations that describe how electrons behave at the atomic scale. Such simulations must handle hundreds or even thousands of atoms simultaneously and carefully track changes in electronic charge, making them extremely demanding in terms of computation and memory. Supercomputing infrastructures enable this by performing numerous calculations in parallel and storing vast amounts of data. Thanks to that, researchers can build realistic models of catalytic materials and uncover the key factors that control their activity – insights that would be impossible to obtain with simpler calculations.

Based on that, it is entirely understandable that the project “Co3O4|CeO2 heterostructure electronic structure role in small gaseous molecules catalysis” was run on the LUMI supercomputer. We spoke with Prof. Filip Zasada, the Principal Investigator of the project, and Leszek Nowakowski, the team leader of the computations from the Jagiellonian University of Krakow in Poland.

How could you synthetically characterise the subject of the research?

The project focuses on understanding how the molecular and electronic structure of cobalt spinel-ceria heterojunctions (Co₃O₄|CeO₂) governs their redox catalytic activity in reactions involving small gaseous molecules such as N₂O and O₂. Using advanced quantum chemical modeling, it aims to describe the atomic and electronic structure of active sites, characterize charge transfer across the interface, and establish structure-activity relationships. This approach will reveal the fundamental factors that control the catalytic performance of Co₃O₄|CeO₂ heterojunctions, providing a molecular-level understanding of their redox behavior.

Could you share some preliminary results of the project run on LUMI?

We are pleased to report that the computational results from our project, executed on the LUMI infrastructure, have been highly satisfactory and have already contributed to peer-reviewed work. In particular:

• In the publication “N₂O Decomposition on Singly and Doubly (K and Li)-Doped Co₃O₄ Nanocubes – Establishing Key Factors Governing Redox Behavior of Catalysts by Leszek Nowakowski, Camillo Hudy, Filip Zasada, Joanna Gryboś, Witold Piskorz, Anna Wach, Yves Kayser, Jakub Szlachetko & Zbigniew Sojka (J. Am. Chem. Soc., 2024, DOI: 10.1021/jacs.4c06587 American Chemical Society Publications ) the HSE06-DFT calculations performed on LUMI enabled detailed analyses of how doping (Li, K, and Li+K) of Co₃O₄ nanocubes influences the work‑function, density of states (DOS), frontier molecular orbital alignment (e.g., 3πₙ(N₂O)–α‑3d z² and 2πₙ(O₂⁻)–α‑3d z²) and interfacial electron‐transfer steps. These calculations required high-throughput runs of large surface models, which the LUMI-C nodes and short queue times facilitated efficiently.
• In the publication “Orbital Resolution of the Reconstruction of CeO₂ (100) Facet–Hybrid‑DFT and COHP Investigations Supported by HR‑TEM Imaging” by Leszek Nowakowski, Filip Zasada, Joanna Gryboś & Zbigniew Sojka (J. Phys. Chem. C, 2025, DOI: 10.1021/acs.jpcc.4c06854 American Chemical Society Publications) the hybrid‑DFT and COHP (Crystal Orbital Hamilton Population) studies of reconstructed CeO₂(100) surfaces relied on LUMI-C nodes to handle detailed orbital‑resolved analyses and large supercells modelling surface reconstructions.

Together, these results demonstrate that the LUMI allocations enabled us to achieve state-of-the-art DFT and hybrid-DFT simulations of complex surface and interface systems, thereby advancing our understanding of structure–activity relationships for heterojunction catalysts such as Co₃O₄|CeO₂.

What prompted you to carry out research with the use of the LUMI infrastructure? How did it contribute to implementing the part of your research related to calculation and analysis?

The use of the LUMI infrastructure was prompted by the need for large-scale quantum chemical (DFT) simulations required to accurately describe the electronic structure and charge transfer processes in Co₃O₄|CeO₂ heterojunctions and their components. The exceptional computational power and parallel efficiency of LUMI allowed us to model complex catalytic interfaces at an unprecedented level of detail, greatly enhancing both accuracy and productivity.

The LUMI resources significantly accelerated our DFT calculations. Although the quantum-chemical software VASP cannot yet utilize LUMI-GPU nodes equipped with AMD Instinct MI250 accelerators, the excellent performance of the dual AMD EPYC 7763 processors, combined with a large number of available nodes and short queue times, substantially improved computational throughput. Additionally, the LUMI-D partition, with dedicated high-memory nodes offering up to 4 TB RAM, enabled the training of Machine Learning Force Fields – a task impossible to perform on standard LUMI-C nodes. Finally, the visualization nodes equipped with NVIDIA A40 GPUs, combined with a fast and easy-to-use web graphical interface, enabled rapid and efficient visualization and analysis of computational data.

How do you evaluate the general performance of LUMI?

We evaluate it very positively. System availability has been excellent, with minimal downtime, and the overall stability of the infrastructure allowed uninterrupted progress of simulations. The availability of pre-installed scientific applications, optimized compilers, and libraries greatly facilitated setting up and running our workflows without significant technical difficulties. The high-performance nodes and fast interconnects ensured that results were obtained much more quickly than on smaller clusters. Additionally, the convenient OpenOnDemand graphical interface, along with excellent and well-documented user guides, significantly improved usability and reduced the learning curve for new users. Overall, LUMI’s combination of computational power, accessibility, and software readiness greatly enhanced both the efficiency and reliability of our research.

Also, the cooperation with the LUMI user support team was excellent. They were very responsive, knowledgeable, and very helpful in resolving technical issues and providing guidance on optimizing our workflows. Their support greatly facilitated the use of the infrastructure.

What further challenges do you foresee in developing research on heterojunctions, and how can high-performance computing help address them?

Further challenges include accurately modeling complex electronic correlations, simulating larger and more realistic systems, and exploring long timescales for dynamic processes. High-performance computing can help address these challenges by providing the computational power necessary for advanced quantum-chemical simulations, enabling the parallel execution of large-scale calculations, and offering access to high-memory nodes for methods that require substantial memory. Additionally, HPC resources facilitate the training and application of machine-learning models to predict properties and behaviors of heterojunctions, which would be infeasible on smaller computing infrastructures.

The team also shared with us the following stages of their research, which involve:

• Extending the DFT modelling to full Co₃O₄|CeO₂ heterojunction, including explicit interface charge‐transfer and defect/dopant distributions.
• Further training machine‑learning force fields (MLFFs) based on the high‑fidelity DFT data generated so far, enabled by LUMI‑D nodes with 4 TB RAM.
• Running reaction‐pathway (NEB) simulations for redox processes (e.g., 2 N₂O → 2 N₂ + O₂) on selected surfaces/interfaces to link molecular insight with catalytic turnover metrics.
• Validating simulation predictions through collaboration with experimental partners for targeted surface synthesis and in‑situ characterisation.

We congratulate the research team on the  results and wish you continued success in your future research endeavors!

Author: Kamil Mucha, Cyfronet