Scientific research and development have transformed and immeasurably improved the human condition, whether by building instruments to unveil the mysteries of the universe, developing treatments to fight cancer or improving our understanding of the human genome. Yet, science can, and frequently does, impact the environment, and the magnitude of these impacts is not always well understood. Given the connection between climate change and human health, it is becoming increasingly apparent to biomedical researchers in particular, as well as their funders, that the environmental effects of research should be taken into account1,2,3,4,5.

Recent studies have begun to elucidate the environmental impacts of scientific research, with an initial focus on scientific conferences and experimental laboratories6. The 2019 Fall Meeting of the American Geophysical Union was estimated to emit 80,000 metric tonnes of CO2 equivalent (tCO2e), equivalent to the average weekly emissions of the city of Edinburgh, UK7 (CO2e, or CO2-equivalent, summarizes the global warming impacts of a range of greenhouse gases (GHGs) and is the standard metric for carbon footprints, although its accuracy is sometimes debated8) The annual meeting of the Society for Neuroscience was estimated to emit 22,000 tCO2e, approximately the annual carbon footprint of 1,000 medium-sized laboratories9. The life-cycle impact (including construction and usage) of university buildings has been estimated at ~0.125 tCO2e m−2 yr−1 (ref. 10), and the yearly carbon footprint of a typical life-science laboratory at ~20 tCO2e (ref. 9). The Laboratory Efficiency Assessment Framework (LEAF) is a widely adopted standard to monitor and reduce the carbon footprint of laboratory-based research11. Other recent frameworks can help to raise awareness: GES 1point512 provides an open-source tool to estimate the carbon footprint of research laboratories and covers buildings, procurement, commuting and travel, and the Environmental Responsibility 5-R Framework provides guidelines for ecologically conscious research13.

With the increasing scale of high-performance and cloud computing, the computational sciences are susceptible to having silent and unintended environmental impacts. The sector of information and communication technologies (ICT) was responsible for between 1.8% and 2.8% of global GHG emissions in 202014—more than aviation (1.9%15)—and, if unchecked, the ICT carbon footprint could grow exponentially in coming years14. Although the environmental impact of experimental ‘wet’ laboratories is more immediately obvious, with their large pieces of equipment and high plastic and reagent usage, the impact of algorithms is less clear and often underestimated. The risks of seeking performance at any cost and the importance of considering energy usage and sustainability when developing new hardware for high-performance computing (HPC) was raised as early as 200716. Since then, continuous improvements have been made by developing new hardware, building lower-energy data centers and implementing more efficient HPC systems17,18. However, it is only in the past five years that these concerns have reached HPC users, in particular researchers. Notably, the field of artificial intelligence (AI) has first taken note of its environmental impacts, in particular those of the very large language models developed19,20,21,22,23. It is unclear, however, to what extent this has led the field towards more sustainable research practices. A small number of studies have also been performed in other fields, including bioinformatics24, astronomy and astrophysics25,26,27,28, particle physics29, neuroscience30 and computational social sciences31. Health data science is starting to address the subject, but a recent systematic review found only 25 publications in the field over the past 12 years32. In addition to the environmental effects of electricity usage, manufacturing and disposal of hardware, there are also concerns around data centers’ water usage and land footprint33. Notably, computational science, in particular AI, has the potential to help fight climate change, for example, by improving the efficiency of wind farms, by facilitating low-carbon urban mobility and by better understanding and anticipating severe weather events34.

In this Perspective we highlight the nascent field of environmentally sustainable computational science (ESCS)—what we have learned from the research so far, and what scientists can do to mitigate their environmental impacts. In doing so, we present GREENER (Governance, Responsibility, Estimation, Energy and embodied impacts, New collaborations, Education and Research; Fig. 1), a set of principles for how the computational science community could lead the way in sustainable research practices, maximizing computational science’s benefit to both humanity and the environment.

Fig. 1: GREENER principles for ESCS.
figure 1

The GREENER principles enable cultural change (blue arrows), which in turn facilitates their implementation (green arrows) and triggers a virtuous circle.

Environmental impacts of the computational sciences

The past three years have seen increased concerns regarding the carbon footprint of computations, and only recently have tools21,35,36,37 and guidelines38 been widely available to computational scientists to allow them to estimate their carbon footprint and be more environmentally sustainable.

Most calculators that estimate the carbon footprint of computations are targeted at machine learning tasks and so are primarily suited to Python pipelines, graphics processing units (GPUs) and/or cloud computing36,37,39,40. Python libraries have the benefit of integrating well into machine learning pipelines or online calculators for cloud GPUs21,41. Recently, a flexible online tool, the Green Algorithms calculator35, enabled the estimation of the carbon footprint for nearly any computational task, empowering sustainability metrics across fields, hardware, computing platforms and locations.

Some publications, such as ref. 38, have listed simple actions that computational scientists can take regarding their environmental impact, including estimating the carbon footprint of running algorithms, both a posteriori to acknowledge the impact of a project and before starting as part of a cost–benefit analysis. A 2020 report from The Royal Society formalizes this with the notion of ‘energy proportionality’, meaning the environmental impacts of an innovation must be outweighed by its environmental or societal benefits34. It is also important to minimize electronic waste by keeping devices for longer and using second-hand hardware when possible. A 2021 report by the World Health Organization42 warns of the dramatic effect of e-waste on population health, particularly children. The unregulated informal recycling industry, which handles more than 80% of the 53 million tonnes of e-waste, causes a high level of water, soil and air pollution, often in low- and middle-income countries43. The up to 56 million informal waste workers are also exposed to hazardous chemicals such as heavy metals and persistent organic pollutants42. Scientists can also choose energy-efficient hardware and computing facilities, while favoring those powered by green energy. Writing efficient code can substantially reduce the carbon footprint as well, and this can be done alongside making hardware requirements and carbon footprints clear when releasing new software. The Green Software Foundation ( promotes carbon-aware coding to reduce the operational carbon footprint of the softwares used in all aspects of society. There is, however, a rebound effect to making algorithms and hardware more efficient: instead of reducing computing usage, increased efficiency encourages more analyses to be performed, which leads to a revaluation of the cost–benefit but often results in increased carbon footprints. The rebound effect is a key example of why research practice should adapt to technological advances so that they lead to carbon footprint reductions.

GREENER computational science

ESCS is an emerging field, but one that is of rapidly increasing importance given the climate crisis. In the following, our proposed set of principles (Fig. 1) outlines the main axes where progress is needed, where opportunities lie and where we believe efforts should be concentrated.

Governance and responsibility

Everyone involved in computational science has a role to play in making the field more sustainable, and many do already, from grassroots movements to large institutions. Individual and institutional responsibility is a necessary step to ensure transparency and reduction of GHG emission. Here we highlight key stakeholders alongside existing initiatives and future opportunities for involvement.

Grassroots initiatives led by graduate students, early career researchers and laboratory technicians have shown great success in tackling the carbon footprint of laboratory work, including Green Labs Netherlands44, the Nottingham Technical Sustainability Working Group or the Digital Humanities Climate Coalition45. International coalitions such as the Sustainable Research (SuRe) Symposium, initially set up for wet laboratories, have started to address the impact of computing as well. IT teams in HPC centers are naturally key, both in terms of training and ensuring that the appropriate information is logged so that scientists can follow the carbon footprints of their work. Principal investigators can encourage their teams to think about this issue and provide access to suitable training when needed.

Simultaneously, top–down approaches are needed, with funding bodies and journals occupying key positions in both incentivizing carbon-footprint reduction and in promoting transparency. Funding bodies can directly influence the researchers they fund and those applying for funding via their funding policies. They can require estimates of carbon footprints to be included in funding applications as part of ‘environmental impacts statements’. Many funding bodies include sustainability in their guidelines already; see, for example, the UK’s NIHR carbon reduction guidelines1, the brief mention of the environment in UKRI’s terms and conditions46, and the Wellcome Trust’s carbon-offsetting travel policy47.

Although these are important first steps, bolder action is needed to meet the urgency of climate change. For example, UKRI’s digital research infrastructure scoping project48, which seeks to provide a roadmap to net zero for its digital infrastructure, sends a clear message that sustainable research includes minimizing the GHG emissions from computation. The project not only raises awareness but will hopefully result in reductions in GHG emissions.

Large research institutes are key to managing and expanding centralized data infrastructures and trusted research environments (TREs). For example, EMBL’s European Bioinformatics Institute manages more than 40 data resources49, including AlphaFold DB50, which contains over 200,000,000 predicted protein structures that can be searched, browsed and retrieved according to the FAIR principles (findable, accessible, interoperable, reusable)51. As a consequence, researchers do not need to run the carbon-intensive AlphaFold algorithm for themselves and instead can just query the database. AlphaFold DB was queried programmatically over 700 million times and the web page was accessed 2.4 million times between August 2021 and October 2022. Institutions also have a role in making procurement decisions carefully, taking into account both the manufacturing and operational footprint of hardware purchases. This is critical, as the lifetime footprint of a computational facility is largely determined by the date it is purchased. Facilities could also better balance investment decisions, with a focus on attracting staff based on sustainable and efficient working environments, rather than high-powered hardware52.

However, increases in the efficiencies of digital technology alone are unlikely to prove sufficient in ensuring sustainable resource use53. Alongside these investments, funding bodies should support a shift towards more positive, inclusive and green research cultures, recognizing that more data or bigger models do not always translate into greater insights and that a ‘fit for purpose’ approach can ultimately be more efficient. Organizations such as Health Data Research UK and the UK Health Data Research Alliance have a key convening role in ensuring that awareness is raised around the climate impact of both infrastructure investment and computational methods.

Journals may incentivize authors to acknowledge and indeed estimate the carbon footprint of the work presented. Some authors already do this voluntarily (for example, refs. 54,55,56,57,58,59), mostly in bioinformatics and machine learning so far, but there is potential to expand it to other areas of computational science. In some instances, showing that a new tool is greener can be an argument in support of a new method60.

International societies in charge of organizing annual conferences may help scientists reduce the carbon footprint of presenting their work by offering hybrid options. The COVID-19 pandemic boosted virtual and hybrid meetings, which have a lower carbon footprint while increasing access and diversity7,61. Burtscher and colleagues found that running the annual meeting of the European Astronomical Society online emitted >3,000-fold less CO2e than the in-person meeting (0.582 tCO2e compared to 1,855 tCO2e)25. Institutions are starting to tackle this; for example, the University of Cambridge has released new travel guidelines encouraging virtual meetings whenever feasible and restricting flights to essential travel, while also acknowledging that different career stages have different needs62.

Industry partners will also need to be part of the discussion. Acknowledging and reducing computing environmental impact comes with added challenges in industry, such as shareholder interests and/or public relations. While the EU has backed some initiatives helping ICT-reliant companies to address their carbon footprint, such as, other major stakeholders have expressed skepticism regarding the environmental issues of machine learning models63,64. Although challenging, tech industry engagement and inclusion is nevertheless essential for tackling GHG emissions.

Estimate and report the energy consumption of algorithms

Estimating and monitoring the carbon footprint of computations is an essential step towards sustainable research as it identifies inefficiencies and opportunities for improvement. User-level metrics are crucial to understanding environmental impacts and promoting personal responsibility. In some HPC situations, particularly in academia, the financial cost of running computations is negligible and scientists may have the impression of unlimited and inconsequential computing capacity. Quantifying the carbon footprint of individual projects helps raise awareness of the true costs of research.

Although progress has been made in estimating energy usage and carbon footprints over the past few years, there are still barriers that prevent the routine estimation of environmental impacts. From task-agnostic, general-purpose calculators35 and task-specific packages36,37,65 to server-side softwares66,67, each estimation tool is a trade-off between ease of use and accuracy. A recent primer68 discusses these different options in more detail and provides recommendations as to which approach fits a particular need.

Regardless of the calculator used, for these tools to work effectively and for scientists to have an accurate representation of their energy consumption, it is important to understand the power management for different components. For example, the power usage of processing cores such as central processing units (CPUs) and GPUs is not a readily available metric; instead, thermal design power (meaning, how much heat the chip can be expected to dissipate in a normal setting) is used. Although an acceptable approximation, it has also been shown to substantially underestimate power usage in some situations69. The efficiency of data centers is measured by the power usage effectiveness (PUE), which quantifies how much energy is needed for non-computing tasks, mainly cooling (efficient data centers have PUEs close to 1). This metric is widely used, with large cloud providers reporting low PUEs (for example, 1.11 for Google70 compared to a global average of 1.5771), but discrepancies in how it is calculated can limit PUE interpretation and thus its impact72,73,74. A standard from the International Organization for Standardization is trying to address this75. Unfortunately, the PUE of a particular data center, whether cloud or institutional, is rarely publicly documented. Thus, an important step is the data science and infrastructure community making both hardware and data centers’ energy consumption metrics available to their users and the public. Ultimately, tackling unnecessary carbon footprints will require transparency34.

Tackling energy and embodied impacts through new collaborations

Minimizing carbon intensity (meaning the carbon footprint of producing electricity) is one of the most immediately impactful ways to reduce GHG emissions. Carbon intensities depend largely on geographical location, with up to three orders of magnitude between the top and bottom performing high-income countries in terms of low carbon energies (from 0.10 gCO2e kWh−1 in Iceland to 770 gCO2e kWh−1 in Australia76). Changing the carbon intensity of a local state or national government is nearly always impractical as it would necessitate protracted campaigns to change energy policies. An alternative is to relocate computations to low-carbon settings and countries, but, depending on the type of facility or the sensitivity of the data, this may not always be possible. New inter-institutional cooperation may open up opportunities to enable access to low-carbon data centers in real time.

It is, however, essential to recognize and account for inequalities between countries in terms of access to green energy sources. International cooperation is key to providing scientists from low- and middle-income countries (LMICs), who frequently only have high-carbon-intensity options available to them, access to low-carbon computing infrastructures for their work. In the longer term, international partnerships between organizations and nations can help build low-carbon computing capacity in LMICs.

Furthermore, the footprint of user devices should not be forgotten. In one estimate, the energy footprint of streaming a video to a laptop is mainly on the laptop (72%), with 23% used in transmission and a mere 5% at the data center77. Zero clients (user devices with no compute or storage capacity) can be used in some research use cases and drastically reduce the client-side footprint78.

It can be tempting to reduce the environmental impacts of computing to electricity needs, as these are the easiest ones to estimate. However, water usage, ecological impacts and embodied carbon footprints from manufacturing should also be addressed. For example, for personal hardware, such as laptops, 70–80% of the life-cycle impact of these devices comes from manufacturing only79, as it involves mining raw materials and assembling the different components, which require water and energy. Moreover, manufacturing often takes place in countries that have a higher carbon intensity for power generation and a slower transition to zero-carbon power80. Currently, hardware renewal policies, either for work computers or servers in data centers, are often closely dependent on warranties and financial costs, with environmental costs rarely considered. For hardware used in data centers, regular updates may be both financially and environmentally friendly, as efficiency gains may offset manufacturing impacts. Estimating these environmental impacts will allow HPC teams to know for sure. Reconditioned and remanufactured laptops and servers are available, but growth of this sector is currently limited by negative consumer perception81. Major suppliers of hardware are making substantial commitments, such as 100% renewable energy supply by 203082 or net zero by 205083.

Another key consideration is data storage. Scientific datasets are now measured in petabytes (PB). In genomics, the popular UK Biobank cohort84 is expected to reach 15 PB by 202585, and the first image of a black hole required the collection of 5 PB of data86. The carbon footprint of storing data depends on numerous factors, but based on some manufacturers’ estimations, the order of magnitude of the life-cycle footprint of storing 1 TB of data for a year is ~10 kg CO2e (refs. 87,88). This issue is exacerbated by the duplication of such datasets in order for each institution, and sometimes each research group, to have a copy. Centralized and collaborative computing resources (such as TREs) holding both data and computing hardware may help alleviate redundant resources. TRE efforts in the UK span both health (for example, NHS Digital89) and administrative data (for example, the SAIL databank on the UK Secure Research Platform90 and the Office for National Statistics Secure Research Service91). Large (hyperscale) data centers are expected to be more energy-efficient92, but they may also encourage unnecessary increases in the scale of computing (rebound effect).

The importance of dedicated education and research efforts for ESCS

Education is essential to raise awareness with different stakeholders. In lieu of incorporating some aspects into more formal undergraduate programs, integrating sustainability into computational training courses is a tangible first step toward reducing carbon footprints. An example is the ‘Green Computing’ Workshop on Education at the 2022 conference on Intelligent Systems for Molecular Biology.

Investing in research that will catalyze innovation in the field of ESCS is a crucial role for funders and institutions to play. Although global data centers’ workloads have increased more than sixfold between 2010 and 2018, their total electricity usage has been approximately stable due to the use of power-efficient hardware93, but environmentally sustainable investments will be needed to perpetuate this trend. Initiatives like Wellcome’s Research Sustainability project94, which look to highlight key gaps where investment could deliver the next generation of ESCS tools and technology, are key to ensuring that growth in energy demand beyond current efficiency trends can be managed in a sustainable way. Similarly, the UKRI Data and Analytics Research Environments UK program (DARE UK) needs to ensure that sustainability is a key evaluation criterion for funding and infrastructure investments for the next generation of TREs.

Recent studies found that the most widely used programming languages in research, such as R and Python95, tend to be the least energy-efficient ones96,97, and, although it is unlikely that forcing the community to switch to more efficient languages would benefit the environment in the short term (due to inefficient coding for example), this highlights the importance of having trained research software engineers within research groups to ensure that the algorithms used are efficiently implemented. There is also scope to use current tools more efficiently by better understanding and monitoring how coding choices impact carbon footprints. Algorithms also come with high memory requirements, sometimes using more energy than processors98. Unfortunately, memory power usage remains poorly optimized, as speed of access is almost always favored over energy efficiency99. Providing users and software engineers with the flexibility to opt for energy efficiency would present an opportunity for a reduction in GHG emissions100,101.

Cultural change

In parallel to the technological reductions in energy usage and carbon footprints, research practices will also need to change to avoid rebound effects38. Similar to the aviation industry, there is a tendency to count on technology to solve sustainability concerns without having to change usage102 (that is, waiting on computing to become zero-carbon rather than acting on how we use it). Cultural change in the computing community to reconsider how we think about computing costs will be necessary. Research strategies at all levels will need to consider environmental impacts and corresponding approaches to carbon footprint minimization. The upcoming extension of the LEAF standard for computational laboratories will provide researchers with tangible tools to do so. Day to day, there is a need to solve trade-offs between the speed of computation, accuracy and GHG emissions, keeping in mind the goal of GHG reduction. These changes in scientific practices are challenging, but, importantly, there are synergies between open computational science and green computing103. For example, making code, data and models FAIR so that other scientists avoid unnecessary computations can increase the reach and impact of a project. FAIR practices can result in highly efficient code implementations, reduce the need to retrain models, and reduce unnecessary data generation/storage, thus reducing the overall carbon footprint. As a result, green computing and FAIR practices may both stimulate innovation and reduce financial costs.

Moreover, computational science has downstream effects on carbon footprints in other areas. In the biomedical sciences, developments in machine learning and computer vision impact the speed and scale of medical imaging processing. Discoveries in health data science make their way to clinicians and patients through, for example, connected devices. In each of these cases and many others, environmental impacts propagate through the whole digital health sector32. Yet, here too synergies exist. In many cases, such as telemedicine, there may be a net benefit in terms of both carbon and patient care, provided that all impacts have been carefully accounted for. These questions are beginning to be tackled in medicine, such as assessments of the environmental impact of telehealth104 or studies into ways to sustainably handle large volumes of medical imaging data105. For the latter, NHS Digital (the UK’s national provider of information, data and IT systems for health and social care) has released guidelines to this effect106. Outside the biomedical field, there are immense but, so far, unrealized opportunities for similar efforts.


The computational sciences have an opportunity to lead the way in sustainability, which may be achieved through the GREENER principles for ESCS (Fig. 1): Governance, Responsibility, Estimation, Energy and embodied impacts, New collaborations, Education and Research. This will require more transparency on environmental impacts. Although some tools already exist to estimate carbon footprints, more specialized ones will be needed alongside a clearer understanding of the carbon footprint of hardware and facilities, as well as more systematic monitoring and acknowledgment of carbon footprints. Measurement is a first step, followed by a reduction in GHG emissions. This can be achieved with better training and sensible policies for renewing hardware and storing data. Cooperation, open science and equitable access to low-carbon computing facilities will also be crucial107. Computing practices will need to adapt to include carbon footprints in cost–benefit analyses, as well as consider the environmental impacts of downstream applications. The development of sustainable solutions will need particularly careful consideration, as they frequently have the least benefit for populations, often in LMICs, who suffer the most from climate change22,108. All stakeholders have a role to play, from funding bodies, journals and institutions to HPC teams and early career researchers. There is now a window of time and an immense opportunity to transform computational science into an exemplar of broad societal impact and sustainability.