About Energy and LCA

MarcLeobet

Use case : LCA of a LLM.
We have made an first step with only Climate change, energy and water : https://hal.science/hal-05685132v1
The final report will assume particules, irradiation & abiotic (ADPe) following AFNOR Spec for frugal AI.
Unfortunately, if nuclear energy production is well documented, the other facilities are less robust.
I understood from your Inferences model that you compute energy production impact - at least in France - taking into account hydro-, solar- wind- and may be gaz and coal. Have you did?

I did not found relevant points in your documentation, and UNECE 2021 LCA of electricity generation option seems to not consider PEF 3.0, which is recommended in UE and bases your work ACV de GPU pour l'IA.

As far as possible, I try to fit to existant methods and I would be glad to add a stone to the wall.

bpetit

Hi,

We do compute the energy production impact. This still has to be better documented, but today it is based mainly on old impact factors coming from Ademe Base Empreinte (except for WU, coming from WRI), as those factors are the only open access factors we found so far for multi criteria electricity generation footprint. You'll find them here: https://gitlab.com/hubblo/dcfootprint/-/blob/main/src/data/materials_impacts.json?ref_type=heads

Next step on this will be to build new factors for electricity, based on known impact factors per electricity generation production source and available, more recent, databases on the mix available in each country.