ATLAS can now filter muon data 15% faster — and may need to, after 2030

Physicists at the Large Hadron Collider built machine-learning systems to process particle collision data faster as the detector gets noisier. One approach speeds up existing reconstruction by 15%; another processes the data end-to-end in 2.3 milliseconds on standard GPUs, which matters because the collision rate will quadruple after 2030.

The High-Luminosity LHC upgrade after 2030 will produce four times as many collisions per second as today. The ATLAS detector's trigger system — the real-time filter that decides which collisions to keep and which to discard — will have to make that decision faster or lose data. This paper shows that standard machine-learning architectures (Vision Transformers and Graph Neural Networks) can do the job on consumer hardware, which means the detector doesn't need a complete rebuild. The constraint was computational speed; the constraint just loosened.