Orateur
Description
What does it take for a model trained on one accelerator to be of any use on the next one? Machines trip, drift, take too long to simulate and cost a lot of energy, and machine learning has answers for each of these. Yet most of those answers stay where they were born: on one machine, with its own data, at the proof-of-concept stage. The ARTIFACT network (41 partners, 14 countries) decided to put the problem on paper. The result is a strategy in seven pillars and a 2026-2031 roadmap, in which data can stay at home while models travel, and every milestone comes with a threshold we can miss.
But paper is patient, and accelerators are not. So the second half of this talk goes to the machines. iRIS is two weeks old and asks how much energy and material AI can save, with HL-LHC, GANIL-SPIRAL2 and FAIR as testbeds. TwinRISE starts in January and will train digital twins across facilities that do not share their data, with its trust criteria fixed before the first model exists. Neither has a result to show yet, and we will say so. What we do have is one week in a manor house in Normandy, where thirty people in six teams ran a single pipeline on data from four machines. We will tell where all this stands, what is still missing, and where physicists who do machine learning can step in.