Reproduce “PELICAN: Permutation Equivariant and Lorentz Invariant or Covariant Aggregator Network for Particle Physics” on OpenResearch

About this paper

Alexander Bogatskiy, Timothy Hoffman, David W. Miller, Jan T. Offermann

PELICAN, a neural network developed by researchers from the Flatiron Institute and the University of Chicago, integrates Lorentz invariance and permutation equivariance for particle physics tasks by using pairwise Lorentz invariant dot products as inputs. This architecture achieved state-of-the-art results in top quark jet classification with significantly fewer parameters and demonstrated superior W boson 4-momentum reconstruction compared to non-ML baselines.

Key insights

• Explicitly embedding fundamental physical symmetries into neural network design yields superior performance and significantly reduced model complexity, achieving state-of-the-art results with substantially fewer parameters.
• Using only the complete set of pairwise Lorentz invariant dot products as inputs is sufficient for effective learning in particle physics tasks, capturing necessary information while simplifying input representation.
• The combination of permutation-equivariant aggregation layers with Lorentz invariant inputs enables efficient and robust modeling of complex particle interactions, supporting both classification and regression.

Summary by alphaXiv.