SuperCode

SuperCode develops a generic AI-assisted co-design methodology for data-intensive science, with radio astronomy as its principal application and validation area. The approach addresses the difficulty of adapting scientific software to new and potentially more energy-efficient computing hardware. In the demonstrated astronomy setting, instruments such as LOFAR and SKA depend on computationally intensive signal-processing pipelines that transform telescope observations into scientific data products. SuperCode uses language models (LLMs) to generate, adapt and optimise scientific code for different hardware architectures, drawing on existing scientific software and reference implementations of operations such as FFTs, beamforming, correlation and dedispersion. More recent work implements iterative code-generation workflows in which smaller language models receive compiler, interpreter and test feedback and revise their outputs. Although current demonstrations focus on radio astronomy, the methodology is designed for wider use in data-intensive scientific computing.

SuperCode was established in 2025 as a public-private research collaboration centred on Leiden University and ASTRON, with participation from scientific-computing infrastructure providers and industrial software partners. The project combines expertise in domain science, high-performance computing, software engineering and artificial intelligence. This interdisciplinary structure underpins a co-design process in which scientific requirements, software, hardware characteristics and AI-assisted development are considered together. Yearly workshops and hackathons support this interdisciplinary collaboration. The practice depends on access to existing scientific code, reference and hand-optimised implementations, heterogeneous computing hardware and the capacity to benchmark generated code for correctness, performance and resource use. The main AI inputs are therefore scientific software, code repositories, technical documentation and reference implementations of radio-astronomy signal-processing algorithms, rather than observational research data. In the astronomy use cases, open-source radio-astronomy software provides both domain context and reference implementations. The project further uses coding language models, retrieval-augmented generation, compiler/interpreter and test feedback, and a public superRAG implementation; repositories are available at https://github.com/SuperCode-Leiden-University/. The work relies on heterogeneous accelerator and HPC systems, including Nvidia and AMD GPUs. Infrastructure such as SURF’s innovation facilities and DAS-6 or successor systems supports access to emerging architectures. Existing foundation models are adapted rather than large models being trained from scratch, using approaches including parameter-efficient fine-tuning, retrieval-augmented generation and iterative tool feedback. Small language models are also adapted and evaluated, and recent work tests multi-sampling and agentic compiler-feedback strategies to improve smaller coding models. Human expertise remains integral: the project describes the approach as semi-supervised and keeps experts within the co-design loop.

SuperCode contributes a practical approach for making AI-assisted scientific software development more resource-conscious while maintaining technical quality. Its scientific relevance lies in supporting the optimisation and porting of software used in data-intensive research, where computing capacity can constrain what analyses are feasible. In radio astronomy, this is particularly important for LOFAR, whose upgrade is expected to increase computational requirements by around forty-fold. Recent experiments show that smaller coding models can be strengthened without automatically moving to much larger models. Under a matched generation-time budget, a 7B model using multiple candidate generations achieved comparable or better results than a 32B model while using one A100 GPU instead of four. An agentic workflow that returned compiler feedback to the model produced further, consistent improvements across the models tested, while executable checks helped maintain code quality. The 2026 study is openly available and the associated superRAG implementation has been released publicly. AI use is explicitly documented, including where models enter the workflow and how compiler, test and retrieval tools are used to constrain and improve outputs; human expertise remains part of the co-design process. The methodology is intended to transfer beyond astronomy to other data-intensive research settings, although such cross-domain application has not yet been demonstrated.

Frugality is a central design principle. SuperCode treats sustainability as a performance criterion for scientific software and hardware co-design, while also testing whether the AI component itself can be made more resource-efficient through smaller models, multiple candidate generations and iterative tool feedback. The current evidence demonstrates lower GPU requirements and computational demand for the AI-assisted development process. The broader project aims to assess scientific computing in terms of energy use and other environmental impacts, but direct reductions in energy consumption, carbon emissions or water use have not yet been quantified in the published experimental results.

P. Chris Broekema, Rob V. van Nieuwpoort: SuperCode: Sustainability PER AI-driven Co-design. The 22nd ACM International Conference on Computing Frontiers, Cagliari, Italy, May 28–30, 2025. pages 141–149, DOI: 10.1145/3706594.3727576.

Elisa Chiarotto, Jingbo Li, P. Chris Broekema, Rob V. van Nieuwpoort: Enhancing SLMs for Sustainable Code Optimization in Radio-Astronomy. arXiv:2607.21677, 2026. https://arxiv.org/abs/2607.21677

Code repository: https://github.com/SuperCode-Leiden-University/superRAG/