Accelerating lithium battery research

Research background

Advanced battery materials are critical for the global transition to renewable energy. At the heart of this movement, researcher Joseph Nelson and his research group at Lincoln University are investigating lithium sponges to find new battery electrode materials. This research focuses on understanding the core structural and scaling properties of materials to build better, more efficient batteries.

To explore these complex molecular arrangements, the team relies on advanced materials modeling, running intensive simulations using the CASTEP software package. The team compiled a bespoke, project-specific binary tailored exclusively for Joseph's environment. Individual calculations take up to one hour. However, a large number of calculations need to be performed to explore the landscape of possible chemical element configurations – requiring millions of core hours of computation. Even minor efficiency gains represent significant saved resources for such a large computational allocation.

Figure 1. Parallel efficiency plots showing core scaling benchmarks and efficiency thresholds across multiple test cases. The performance varies significantly with the toolchain used to compile CASTEP.

Project challenges

Joseph came to the team seeking to optimize his group's simulation workflows. The project faced an exceptionally high burn rate of $80,000 NZD per quarter due to computational inefficiencies. They initially relied on the software developer's default recommendation to run CASTEP version 24.1 using 64 cores per job. Jobs typically run across 8 to 128 processor cores, but adding more processors yielded little to no scaling benefit. The team lacked clear benchmarks regarding compiled binaries, academic licensing constraints, and multi-core scaling behaviors on modern CPU nodes.

What was done

REANNZ’s Research Software Engineers mapped out the CASTEP package constraints and executed multi-core CPU scaling tests. The team developed custom scripts to test performance across different core counts. Testing revealed that scaling down from the recommended 64 cores to just 8–16 cores per job unlocked a 4x to 8x improvement in job efficiency, directly addressing the client's high burn rate. 

John Whiting, one of REANNZ's Research Software Engineers who worked on the project at the time, successfully completed the initial coupling, linking the binary to optimized FFT and math packages without major compilation difficulties. The engineers then benchmarked the FlexiBLAS backends at runtime using the FLEXIBLAS environment variable to identify further performance variations. They evaluated performance across OPENBLAS, BLIS, NETLIB, and AOCL_MT using small (4 MPI ranks) and large Li6Ti4O3 (16 MPI ranks) foss/2026 cmake build test cases.

Figure 2. Comparing runtime performance across different core counts and software configurations. Some configurations only scale effectively up to 8-16 cores.

Additionally, the team wrote a standalone FFT test to tease out performance differences between FFT implementations, evaluating FFTW and AOCL-FFT for large problem sizes. Control of these optimization steps was subsequently handed over to the users to increase local ownership of the workflow.

Main outcomes

Together, these optimisations proved transformative: scaling down the core counts achieved a massive efficiency breakthrough, showing that an 80% parallel efficiency threshold holds true across the board. The benchmarking process successfully identified OPENBLAS as the fastest, most optimal backend configuration for their setup, while the FFT tests confirmed a performance tie between FFTW and AOCL-FFT for large problem sizes.

Crucially, the consultancy dramatically reduced the client's $80,000 NZD quarterly computational spend while equipping the research team with the workflow controls to maximize future computing speeds. Looking forward, the team mapped out a clear path for GPU offloading. While current commercial bundles do not support GPU calculations, a public beta release of the academic CASTEP GPU port (based on version 23.1) utilizes OpenACC directives to offload intensive tasks like Fast Fourier Transforms (FFTs) and nonlocal exchange-correlation functionals. Full GPU support is anticipated in version 25.1 to accelerate future material science research.

Researcher feedback

Working alongside the REANNZ team let us unlock large efficiency gains, and make significantly better use of the REANNZ platforms. This accelerates our research into new battery materials, enabling a greater number of prospective battery chemistries to be tested. This increased efficiency reduces the time-to-science for our battery research.

Joseph Nelson, Lincoln University

This case study shares some of the technical details and outcomes provided through our Consultancy Service. This service supports projects across a range of domains, with an aim to lift researchers’ productivity, efficiency, and skills in research computing. Get in touch to discuss how our Research Software Engineers and specialist support could help advance your project.