Nature Machine Intelligence presents EveNet, a shared model for diverse particle-collision analyses
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A single pretrained model can serve as the foundation for diverse physics analyses and reduce the need for large labelled datasets for each new task.
EveNet is pretrained on 500 million simulated collisions, combining self-supervised learning with physics-informed supervision. Instead of a separate model per task, it builds a shared representation of collision-object sets of variable length. In searches for heavy resonances and exotic Higgs boson decays, the model outperformed established and state-of-the-art baselines, especially with few examples. On public CMS data, the researchers tested transfer to experimental data, rediscovering the Υ meson. The pretrained model also enabled accurate extraction of quantum-correlation measures that were robust to the systematic variations tested.
Sources 1