AI for particle physics preserves a key rule of relativity at little extra cost

A neural-network design keeps predictions consistent with Lorentz symmetry while supporting several particle-physics tasks.

Editorial illustration — not from the study.

The method, called Lorentz Local Canonicalization, predicts a local reference frame for each particle in a Large Hadron Collider setting. It then transfers information between those frames while preserving how that information should change under Lorentz transformations, which describe rotations and changes between steadily moving observers.

The researchers tested the design with graph networks and transformers. They report applications to amplitude regression, end-to-end event generation and jet tagging, and introduce a large dataset for benchmarking models that identify top-quark jets.

What the design does

Lorentz Local Canonicalization provides exact Lorentz-equivariance for arbitrary neural networks, according to the paper, with minimal computational overhead. In the particle-physics applications studied, it predicted local reference frames for individual particles and propagated tensorial information of any order between those frames. The researchers incorporated the approach into graph networks and transformers and demonstrated it on amplitude regression, end-to-end event generation and jet tagging. They also introduced a large top-tagging dataset to compare versions of established benchmark models and examined the role of symmetry breaking in jet tagging.

Why relativity matters

Particle-physics data are governed by relativity, so models that respect Lorentz symmetry can build that physical constraint directly into their calculations. The reported design is intended to do this without requiring a major increase in computational cost, while remaining compatible with common neural-network architectures and several types of particle-physics analysis.

Evidence and caveats

This is a journal article reporting methodological tests and applications across several machine-learning tasks. The abstract does not provide numerical results for accuracy, computational overhead or comparisons with individual benchmark models, nor does it describe the size or composition of the new dataset. It therefore supports the reported demonstrations but does not, from the abstract alone, establish that the method is best for every particle-physics task or neural-network architecture.

A neutron-star merger finds no sign that gravity’s strength changed over time

A joint analysis of the neutron-star merger GW170817 and the gamma-ray burst GRB 170817A found no evidence that the gravitational constant has changed over time. The study limits its fractional rate of change to between −3.36 × 10⁻⁹ and 5.34 × 10⁻¹⁰ per year.

1 day ago·9 views

Physics & Space

Two room-temperature lutetium clocks agree at the 19th decimal place

Two room-temperature lutetium-ion clocks achieved fractional frequency uncertainties near one part in 10¹⁹ and agreed in a direct comparison. The result supports efforts to improve international timekeeping and develop a future definition of the second.

2 days ago·11 views

Physics & Space

Physicists refine the map of matter just after the Big Bang

A review of quantum chromodynamics surveys what scientists have learned about the phases of matter that existed about 10 microseconds after the Big Bang. It focuses on particle production, fluctuations and comparisons between collision data and theoretical predictions.

3 days ago·12 views