Cleaning up computing: Sustainability in computational research

Nat: As a Geography and Environmental Science student, I’ve always approached sustainability by tracing the links between environmental, social, and technological systems, but I didn’t expect my studies to lead me into the intersection between environmental research and computer science. This KURF project stood out for its unique combination of sustainability and coding - two areas I’ve grown increasingly passionate about during my undergraduate degree. As I transition into my postgrad studies, I've come to realise more recently how essential it is to recognise the environmental costs of research itself, especially the often hidden impacts of digital tools and computational methods we use every day.

Liz: Just as our mobile phones and laptops are an integral part of our everyday lives, computing has become a crucial part of doing research - whether that’s analysing biomedical datasets, simulating ocean currents, or training an AI model using social media data. The scale of computing required for research often goes beyond what an ordinary laptop can handle, requiring the use of High Performance Computing (HPC) facilities instead. At King’s, HPC facilities are provided by the e-Research department. As a Research Software Engineer working in e-Research, part of my job is working with researchers to help them use these facilities effectively.

Since the foundation of e-Research in 2019, the number of researchers using HPC has grown - and so has the amount of hardware and energy that’s required to power their research! In the context of worldwide climate change, researchers and research funders are becoming increasingly interested in understanding the environmental impacts of research. Therefore, when the opportunity arose to supervise a KURF project, it seemed like a great opportunity to develop tools to help researchers understand these environmental impacts.

Nat: My role as a research fellow was to help adapt an existing tool, developed by ARCHER2, the UK National Supercomputing Service, to estimate the greenhouse gas emissions generated by computational research that uses HPC facilities. More specifically, the tool aims to estimate both the emissions from the electricity used and the “embodied carbon” associated with the research. Embodied carbon refers to all emissions from the manufacturing process of the computing hardware including the extraction of materials, transport and disposal. The first step in this project was to find reliable embodied carbon data for the computer components, which was challenging since most of this information was not readily available. Calculating the emissions from both the use and manufacture of computing hardware offers a more complete picture of the environmental footprint of digital research. While it’s easy to focus on energy usage, the manufacturing process contributes a large chunk of overall emissions (see Figure). The carbon data sourced for this research is the foundation of the tool's emissions calculations.

Liz: The tool that we’ve developed during this project will allow researchers using the HPC facilities at King’s to estimate the carbon emissions associated with their research workflows. It also puts these into context by providing comparisons to the emissions from a typical UK household, or to food production. There’s still more work that can be done to improve the tool - in particular, to get better estimates of the energy usage of individual compute tasks - and I’ll be continuing to work on this in the coming months, together with other colleagues in e-Research. However, the tool is now being tested by a small number of people, and we plan to make this available to all HPC users before the end of the year. We hope that this tool will give researchers information they can use to optimise their computational workflows and prioritise their research tasks. It also forms part of our work towards Green DiSC certification - a new digital sustainability certification scheme for research groups and institutions.

Nat: A major learning from this entire experience was that research is not always straightforward! One of the biggest challenges I encountered was sourcing reliable data, particularly when it came to embodied emissions. Manufacturers often lack transparency around the environmental impact of their hardware, which raised broader concerns for me about the limitations of current sustainability reporting in the tech sector. As conversations around green technology and digital responsibility grow, it becomes increasingly important for institutions and companies to move beyond surface-level sustainability claims and provide transparent, verifiable data. Without access to emissions data, efforts to assess environmental impact risk being incomplete or inaccurate with the data in this project being as accurate as how manufacturers report it to be.

Liz: As a scientist, I always want to have precise and accurate data - but it’s not always available! One thing that I learnt in this project is that it’s better to have some data, even if it’s just an estimate, than to go down a rabbit hole of trying to find the perfect dataset or metric - which might not exist. Even though the outputs from our tool are estimates, they should be in the right ballpark. The important thing is that we’re making information available that wasn’t previously accessible to researchers, and this will allow them to start taking sustainability into account when doing computational research. And of course, we’ll adapt and improve the tool as better data becomes available.

Concluding thoughts (both): Through this project, we combined our data and coding skills and our different perspectives to integrate environmental science and computational methods. More importantly, it has reinforced the belief that while computation is vital to research, including research in climate change and sustainability, sustainability must also address the often-overlooked environmental costs of technology itself. As research, industry, and daily life become ever more data-driven, the invisible carbon costs of computation will continue to rise. From the energy consumed in training AI models, to the power demands of running large-scale scientific simulations, to the resources required for storing and moving datasets, every digital process leaves an environmental mark. If we want technology to be part of the solution, we also have to confront and reduce the footprint of the digital systems we depend on.