Carbon footprint of my PhD

In order to mitigate the effects of climate change and prevent the climate system from reaching tipping points, it has become urgent to drastically reduce greenhouse gas (GHG) emissions¹, and the first step in committing human society to a low-carbon transition is to draw up an inventory of anthropogenic emissions by sector.

During my thesis, I actively participated in a working group set up to estimate the carbon footprint of my laboratory. This post adopts a finer degree of estimation by presenting the individual carbon footprint of my own PhD. Let’s take a look at the main sources of emissions.

Computing

As part of Cerfacs carbon footprint estimation, we have estimated the emission factor of in-house computation to be in the order of 2.2 gCO$_2$eq per core hour of computation. This figure takes into account power and cooling consumption, which are accurately measured, as well as the carbon emissions of the supercomputer life cycle, i.e. including manufacturing, transportation and recycling, which are estimated to be half of the total emissions based on a small literature review. I also asked for the emission factor of the national computer center where I ran the simulations, and if I did not get an answer, I took an average emission factor corresponding to supercomputers of similar size and generation.

Using these emission factors, we can proportionally estimate the GHG emissions of the simulations carried out in this thesis, thanks to an exhaustive inventory of all the base hours calculated with each supercomputer. Note that for some supercomputer, the total amount of core hours consumed can be found a posteriori using the sacct slurm command.

Figure 1: History of core hours computed during my thesis (from 10–03–2021). Each color corresponds to a different supercomputer.

I estimate the total GHG emissions associated with my thesis calculations at around 10 tCO$_2$eq, of which 6 tCO$_2$eq correspond to the ensemble of 200 large-eddy simulations required to build the reduced-order model.

This estimate does not include emissions related to data storage because it requires significantly less energy than computation, and because it is much more difficult to attribute emissions from storage servers to each user. Emissions linked to data transfer can also be considered negligible ².

IT equipment

Using the Ecodiag tool, I estimate the emissions associated with the entire lifecycle of the IT equipment purchased for my thesis:

• 1 MacBook Pro 13" → 256 kgCO$_2$eq,
• 1 Dell P2419HC 24" monitor → 430 kgCO$_2$eq,
• 1 LG HDR UHD 4K 27" monitor → 430 kgCO$_2$eq,

for a total of around 1.1 tCO$_2$eq. It is interesting to note that emissions are directly linked to screen size.

Traveling

For emissions related to business travels, I use the GES 1point5 travel simulator to estimate emissions of each travel I did based on the means of transport:

• 1 field experiment in Lannemezan, by car → 57 kgCO$_2$eq,
• 1 conference in Aveiro, mainly by plane → 444 kgCO$_2$eq,
• 1 lecture series in Brussels, by train → 32 kgCO$_2$eq,

for a total of around 0.5 tCO$_2$eq, the majority of which is due to my return flight to Lisbon. Note that for approximately the same distance, the emissions associated with travelling to Brussels by train are much lower.

Commuting

I also use the GES 1point5 commutes simulator to estimate the emissions associated with my daily commuting: 6km twice a day by bike, 5 days a week. This gives around 18 kgCO$_2$eq after 3 years, which is really negligible compared to the total carbon emissions of my PhD.

Laboratory operation

I also include in my PhD carbon footprint emissions related to the Cerfacs laboratory heating, electricity consumption and cooling/safety fluid leaks. To do so, I use our estimates of the laboratory-level emissions and divide them by the number of people working at Cerfacs over the October 2020 - November 2023 period:

• Heating → 1.3 tCO$_2$eq,
• Electricity consumption → 0.6 tCO$_2$eq,
• Fluid leaks → 0.3 tCO$_2$eq.

Note that emissions linked to supercomputer electricity consumption are not included, as this has already been taken into account in the computation emissions.

---

Conclusion

The total thesis carbon footprint is finally estimated to be of the order of 14.2 tCO$_2$eq. For comparison, 1 tCO$_2$eq represents approximately the emissions associated with a return flight from Paris to New York and 10 tCO$_2$eq the annual footprint of a French person in 2022³. The annual footprint of this thesis is therefore around 5 tCO$_2$eq, which is far more than the 2.3 tCO$_2$eq per capita in France estimated to be required by 2050 to limit global warming to 1.5°C ⁴. It should be noted that these estimates remain orders of magnitude, given the considerable uncertainties involved.

Figure 2: Breakdown of contributions to the thesis carbon footprint.

Computing accounts for by far the largest share of GHG emissions (73%), but these emissions were difficult to limit because simulation was at the heart of the PhD project. This highlights the need to develop efficient sampling and learning methods to limit the carbon footprint of data-driven models and ensemble predictions. Overall, this evaluation exercise raises awareness of the issue of carbon emissions in the research sector and raises questions about the future direction of research.

More information on the carbon footprint of my thesis can be found in Appendix C of my PhD manuscript.