Battery research powered by LUMI

At the Technical University of Denmarks Department for energy conversion and storage (DTU Energy) postdocs Manuel Dillenz and Jose Maria Castillo Robles work to understand how batteries function at the most fundamental level, which is key to improving their performance. This involves tracking how electrons move through materials and how small structural changes affect that movement. Because these processes take place at the scale of individual atoms and unfold in fractions of a second, they are difficult to observe experimentally. Understanding them through simulations requires both highly accurate calculations and sufficient computing power to study realistic systems over the required nanosecond timescales.

“These processes take place on ultrafast timescales and are tightly linked to distortions in the material structure. To capture them accurately, we need highly precise simulations but also the ability to scale up.”, Manuel Dillenz

Small-Scale Science, Large-Scale Computing

To understand what happens inside the material, Manuel and Jose Maria combined several computational approaches which each addressed a different part of the problem. 

First, highly accurate quantum mechanical calculations were carried out on LUMI's CPU partition to understand how charge moves through the material. These calculations provided the reference data needed for the next stages of the project. They then used molecular dynamics simulations to study how the material behaves over time and how its structure changes under different conditions. Together, these simulations generated the data needed to train a machine learning model. Once trained, the model was deployed on LUMI's GPU partition, where it could reproduce the accuracy of the quantum mechanical calculations at a fraction of the computational cost. 

Making the Workflow LUMI-Ready

Before moving to large-scale computing, Manuel and Jose Maria had developed and tested the workflow on the DTU HPC facility Niflheim. This had allowed them to refine the setup and ensure that each step in the process worked as intended. Moving the machine learning workflow from DTU's Niflheim cluster to LUMI was, however, not entirely straightforward. LUMI-G uses a different GPU architecture which means that software optimised for Niflheim’s NVIDIA GPUs could not simply be transferred and run on LUMI's AMD-based GPU infrastructure. To solve this issue, Manuel and Jose Maria reached out to the LUMI user support team through the DeiC Helpdesk. The team recommended a container-based approach, allowing the software and all its dependencies to be packaged into a portable environment. The setup was first tested in an existing software container and later packaged into a dedicated container for the large-scale simulations.

“Adapting our workflow to a new architecture was challenging, but the LUMI support team made the process straightforward. Communication ran through the DeiC Helpdesk over email, and the responses were prompt and knowledgeable. Their recommendation of a container-based setup was exactly the right approach and saved us considerable time”, Manuel Dillenz

Once the software had been adapted to the LUMI environment, access to LUMIs large-scale computing resources made it possible to carry out the most computationally intensive parts of the workflow from generating reference data to training and deploying the machine learning model in large-scale simulations:

A tool for future battery research

While Manuel and Jose Maria's work focused on a specific type of battery material (lithium manganese oxide), the project has produced tools that can be used beyond this single system. This means that researchers working on similar battery materials can reuse both the computational workflow and the trained machine learning models in their own work. The approach can also help explain experimental observations by revealing what is happening at the atomic level inside a material:

“The approach we have developed in this project is not limited to a single material. The workflows and trained machine learning models will be made openly available, which will allow other researchers to build on the work and apply it to related systems.”, Jose Maria Castillo Robles

Looking ahead, the researchers plan to extend the work to different material compositions and to investigate how imperfections in the material structure and small amounts of added elements influence charge transport. Continued access to large-scale computing resources will be important as the research moves towards increasingly complex battery systems.

“I would absolutely apply for LUMI resources again. The combination of CPU and GPU partitions fits this kind of workflow perfectly, and continued access will be essential as we move towards larger and more complex battery systems”. Manuel Dillenz, 

Tools and ressources