The Carbon Footprint of Climate Science – an opinion by Hilary Weller 

By: Hillary Weller

What is the acceptable carbon footprint of climate science? Climate science cannot be done without a carbon footprint, and without climate science we would not know that burning fossil fuels is causing dangerous climate change. So without climate science, the world would burn its way to a largely uninhabitable planet. So surely the carbon footprint of climate science is worth it? I claim the following: 

  1. To make accurate predictions, we need supercomputers that have a carbon footprint equivalent to around 10,000 houses. 
  2. To improve climate predictions, we need to run variations of experimental models on supercomputers.  
  3. To do the best climate science, we must communicate internationally, and communication is best face to face. 
  4. To make progress, early career scientists need to travel widely gain knowledge of the internationally leading edge of science, gain a reputation and to develop a network of collaborators. 

Here comes the “but” … 

But, if the purpose of climate science is to predict the outcomes of a range of emissions scenarios and to inform the policy that will eradicate CO2 emissions, then surely we must do this with a reduced footprint. We are moving in the right direction – taking the train to European meetings, reducing attendance at meetings that require long haul flights and making use of regional hubs so that international meetings can be held on multiple continents simultaneously. But I argue that we must move faster. I believe that climate scientists should lead the way in low emission science. Our communication may be stilted and inefficient as a consequence, and this may slow the progress of our careers and of the science itself. But the cost is too high to keep travelling. I do not believe that we should be telling early career scientists to take long haul flights for the sake of their careers and for the advancement of science. Instead, we should be asking them how we can communicate more sustainably. My son (aged 11) had an active online social life during lockdown. I cannot picture being able to communicate in a relaxed, friendly, casual and productive way online, with chance meetings over a poster and derivations on a napkin at dinner leading to fruitful collaborations. But we need to learn how to do this with the next generation rather than insisting that long haul flights are needed for the widest possible communication of science. 

Back to the supercomputers. A carbon footprint similar to 10,000 houses seems reasonable for making weather predictions that enable the world to make more carbon efficient choices, saving far more than the initial outlay. (The 10,000 houses comparison was based on some quick web search) But there are supercomputers doing research simulations that may never have an impact. Without the research we cannot have the operational weather predictions which are so beneficial. But there doesn’t seem to be much restraint on research computing. Perhaps research grant proposals in all fields should have to estimate and justify their carbon footprint as well as their expenditure. 

This blog has been political rather than a science notebook which is the expectation. So a little now about the science that I do. I do not have high profile or a high impact career. I do, I think, some interesting and novel research that has the potential to improve weather and climate models. I have been doing some work recently about how to take long time steps without leading to spurious oscillations by using implicit time stepping for advection. This is far cheaper than previously thought and does not have much impact on accuracy. If you can increase time steps then you can reach a solution more quickly, using less computer power.  

Cite: “Adaptively implicit MPDATA advection for arbitrary Courant numbers and meshes”. Hilary Weller, James Woodfield, Christian Kühnlein, Piotr K. Smolarkiewicz, 2022. 

I have also done some work on convection parameterisation – a method of representing clouds and precipitation without high spatial resolution. This is old fashioned. More recently, high resolution simulations with fewer parameterisations have led to more realistic simulations. But if we can make parameterisations more realistic, then we can reduce the need for high resolution simulations that need the biggest supercomputers. My work has been more mathematically interesting than impactful (so far). But I would love to see more work on parameterisation to enable realistic simulations at lower resolution and hence smaller footprint. 

Cite, eg: “Two-fluid single-column modelling of Rayleigh–Bénard convection as a step towards multi-fluid modelling of atmospheric convection”. Daniel Shipley, Hilary Weller, Peter A. Clark, William A. McIntyre, 2021. 

Comments from Colleagues 

Pier Luigi Vidale 5/7/23: “We heard from the CEO of NVIDIA this morning. On their new Grace-Hopper based supercomputer, they can run ICON at 2.5km globally for short time scales, and the energy cost of the run, compared to a traditional multicore supercomputer, is 1/250. He claims that this is just the start, and a bit more can be done, but I think that it is already quite impressive. 

Grace-Hopper combines an ARM-type multicore CPU with a modern NVIDIA GPU, with nearly zero latency in terms of IO and memory access.” 

Thorwald Stein 3/7/23: Your two publications hint at ways to reduce the supercomputer carbon footprint in the future. To provide a positive message for ECRs [early career researchers], I wonder if you could include examples of a future for conferencing, too. One of the best conference interactions I ever had was at a video call initiated through Gather.Town and I’m sad that I’ve not seen that platform used much since. My worst conference was “hybrid” where I stayed up at home until midnight to present my poster, but it was scheduled at lunch time for all in-person attendants sad Seeing virtual conferences as the future rather than a temporary necessity for 2020-2022 requires a major culture shift. Taking it to the extreme, if humanity is ever going to colonise space, video conferencing is here to stay:  

Hilary Weller 4/7/23: I like online meetings when there is unmoderated chat so that lots of discussion about talks goes on during the talks. The best online meeting I went to was PDEs on the sphere in 2021 when we had an open Google doc that we all wrote in, discussing the talks. There were also nice break out rooms where we could catch up with old friends and one person there made sure that everyone introduced themselves. We probably need more online ice-breaker events. I agree, and tools like that could be used more. But I think they need to be part of the timetabled day and with posters rather than just evening socialising, when you really want to get away from your computer. I also think that scientists should use online discussion groups more, such as with Slack. 

Anon 3/7/23 commented on Hilary’s statement “climate scientists should lead the way in low emission science”: This is a good point. Some people conflate “environmental scientists” with “environmentalists” which I find odd. Do we have a greater moral responsibility than those outside our field? 

Anon 3/7/23: Covid forced us to investigate better ways to ‘mingle’ online. I don’t think we’re anywhere near there, but it has to be the goal. The next generation, surely, will think nothing of working closely with others in a globally distributed community. Furthermore, I think science is due for a change in culture. I’ve never been a fan of the cult of the individual superstar, probably because I’m not one, but also because so much of today’s science isn’t about one person sitting in a lab or office coming up with a revolutionary idea. Look at e.g. CERN; in our case, no individual can claim to have generated a climate simulation, but if one or two say something profound about one in Nature, they are lauded as great scientific leaders. We left the Enlightenment a while ago. 

Anon 3/7/23 commented on the statement about supercomputers .. “for making weather predictions that enable the world to make more carbon efficient choices”: Of course, this isn’t the main purpose of NWP, with a few exceptions (one of which is routing long-haul flights …)) 

Hilary Weller 6/7/23: There are loads of examples, mostly because saving fuel saves money. Using renewable energy efficiently needs accurate weather forecasts, ships are routed to sail downwind, people walk or cycle to work based on the forecast, gas tankers are sent to regions that are going to be experiencing cold, calm winter conditions, supermarkets reduce foot waste by providing the food we want for a summer barbecue. 

Anon 3/7/23: I agree, many modelling centres are working on best practice guidelines. And CMIP7 preparations include environmental considerations. But it is true that models are also getting costlier, outputting more data. 

Anon 3/7/23: I have a discomfort when it is stated that supercomputers are using 100% renewable energy, which is a possible retort to the points here. My discomfort is that that renewable energy could be used for something else. Perhaps the debate has moved on over this, but I don’t know how this gets factored into discussions on renewable use.  

Anon 3/7/23: There is huge practical constraint on research computing. For example, we do not do a fraction of the hindcasting/re-forecasting we really should do to characterize our models. Whether all the CPU used is justified is certainly questionable, but it is the nature of research not to know the outcome beforehand. We could always make good use of more! But surely the issue here is about source of energy as well as amount. We have made progress in being able to locate supercomputers remotely from users, and energy use is already a major constraint, but should it be higher? 

Liz Stephens 3/7/23: In a recent call with the funder of our new grant they put to us (informally) that we should be aware of the carbon footprint of our experiments when running them, and make sure that they are all useful/necessary. 

Richard Allan 3/7/23: In terms of the IPCC work (which certainly does have impact on climate policy) although initial in person meetings (involving long haul flights) are I think essential in building the relationships necessary to collaborate and in ensuring diversity in contribution from scientists across the world, the pandemic showed that we can work effectively online, including in agreeing the summary for policy makers line by line with hundreds of government representatives.  

Anon 3/7/23 commented on my research on time stepping: “So your impact, potentially, is to substantially reduce the footprint of climate models and/or improve their accuracy.” 

Hilary: Thanks. Yes, I can have an impact if I can persuade other model developers to adopt the approach and if the approach proves useful in more practical settings. 

Pier Luigi Vidale 18/10/23: A couple of comments and clarifications. The first one is: what is the benefit of such HPC simulations for society? 

In other domains, e.g. medicine, material science, fundamental physics, the typical project is currently using far more HPC than weather and climate applications, yet no such questions about the carbon footprint are asked, mostly because it means that in those domains they can give up most of the lab experimentation, with enormous savings (often also with far more ethical protocols, when life is involved) and incredible speedups in developing new medicines, therapies, vaccines, materials, engines for cars and airplanes, etc.. In our domain we do not have a physical lab, and we are right to be asking whether we are consistent when we say that people should reduce their CO2 footprint, but we must also consider what the benefits of our simulations are. 

In most European grants, both for science and for HPC, we must always demonstrate what the societal benefit is. 

In PRIMAVERA we did use a substantial amount of supercomputing, but: 

  1. a) it was more efficient to run 8 GCMs at 25km, versus running a large number of regional models, at the same resolution, albeit without even covering the entire planet. Many groups worldwide run such downscaling experiments, and there is a lot of needless replication. But they are under the radar, because they do not use one large facility.
  2. b) the global capability in PRIMAVERA meant that industries such as the energy industry and the water industry were involved, and work we did with those industries means that they have a much clearer and more applicable estimate of their global risks and opportunities, as well as new data that they can use to manage their business (e.g. for trading renewable energy across the whole of Europe)
  3. c) PRIMAVERA outputs were widely used by the entire international community, still are (actually my 2012 UPSCALE data are still in use for publication to this date), and PRIMAVERA papers were cited 150 times in the IPCC report

In the current projects, NextGEMS and EERIE, we are working with energy (particularly solar in NextGEMS), fisheries, transportation, again to help society improve the way that resources are used. So yes, using supercomputers has a CO2 footprint, but if it helps reduce other footprints generated by other human activities, there is potential to compensate. This should be researched further. 

Before we go to NVIDIA and GraceHopper, important advances in software engineering over the last 15 years have meant that many groups can now use GPUs for their weather and climate simulations. In the COSMO consortium (Austria, Germany, Italy, Switzerland) this has reduced the energy footprint of the models to 30% of the original. ICON, the current weather and climate model used in Germany and Switzerland, has the same capabilities, and is starting to run on LUMI, which is a hybrid machine, with many GPUs. The IFS is undergoing the same technological changes, and so is NEMO. Using 1/3 of the electrical power is perhaps not going to make a substantial difference, but in terms of investments in software it was just the start, and many believe that it is possible to improve this. NVIDIA is helping the ICON developers far more, now that ICON has been ported to hybrid architectures. 

In Euro-HPC we are discussing charging research groups for the KWh, not for the core hours, so that it is up to them to become more efficient if they want to run long simulations or large ensembles. Also, for the UK, do remember that Archer and Archer2 are run entirely on renewable energy. LUMI, one of the three European exascale machines, located in Finland, promises to do something very similar. 

About sdriscoll Researching machine learning and thermodynamics of Arctic sea ice. Part of SASIP (2021-present) @UniofReading (Schmidt Futures). Previously DPhil Physics @UniofOxford (climate/volcanoes/geoengineering). Also nuclear war/winter + X-risk.
This entry was posted in Climate. Bookmark the permalink.

Leave a Reply

Your email address will not be published. Required fields are marked *