Semi-Visible Jets: How Machine Learning Regresses the Invisible Fraction of Dark Matter at Colliders

Forum topic · 小凯 · 2026-05-04

Summary

This post introduces a particle physics paper (arXiv: 2604.20456) on semi-visible jets (SVJs), a proposed collider signature of dark matter. In an SVJ event, a high-energy shower at the LHC contains a mix of Standard Model hadrons and dark-sector particles that escape the detector invisibly. The key model parameter is r_inv, the fraction of dark hadrons that decay into invisible dark matter: r_inv = 0 means fully visible jets, r_inv = 1 means fully invisible, and intermediate values define the semi-visible regime. The paper's core contribution is a machine learning regression approach that reconstructs r_inv from detector observables: using SVJ events tagged with a high-energy photon, the model learns jet substructure—energy flow, momentum, and shape—to estimate the invisible fraction, analogous to inferring a buried half from the exposed half. The author frames this as a Feynman-style inference problem: deducing unobservable entities from indirect evidence such as missing momentum and anomalous jet shapes, and closes with broader lessons for reasoning in partially observable systems.

Background

This post discusses the paper "How Invisible: Regressing The Key Model Parameter for Semi-visible Jet Searches" by Yin Li, Bingxuan Liu, Jianbin Wang, Jiaqi Xie, Kairong Xu, Ruihan Ye, and Zihuan Huang (arXiv: 2604.20456, 2026-04-27).

Dark matter neither emits nor reflects light and interacts only via gravity—yet under extreme conditions at colliders, it may leave detectable traces.

Semi-Visible Jets (SVJs)

At colliders like the LHC, proton collisions could produce dark matter particles, which normally pass through the detector without a signal. Theory predicts a subtler phenomenon: semi-visible jets. When quarks or gluons fragment into collimated sprays of hadrons (jets), an SVJ contains a mixture of ordinary visible hadrons and dark hadrons that decay invisibly—like a fireworks burst where only half the sparks can be seen.

The Key Parameter: r_inv

The fraction of dark hadrons that decay into invisible dark matter is governed by $r_{\mathrm{inv}}$ :

$r_{\mathrm{inv}} = 0$ : fully visible, standard-model-like jets
$r_{\mathrm{inv}} = 1$ : fully invisible decay products
$0 < r_{\mathrm{inv}} < 1$ : the semi-visible regime

Measuring

r_{\mathrm{inv}}

is effectively measuring "how dark the dark matter is."

Machine Learning Approach

The paper's central contribution is a regression model that reconstructs $r_{\mathrm{inv}}$ from detector data:

1. The collider produces SVJ events (accompanied by a high-energy photon) 2. The detector records the visible energy, momentum, and directions 3. A machine learning model analyzes jet substructure and distributions 4. The model outputs an estimate of $r_{\mathrm{inv}}$

The author likens this to inferring the original length of a chopstick broken in two: you see the exposed half but must infer the buried half.

Feynman-Style Inference

Echoing Feynman's demonstration that electron properties can be inferred from cloud-chamber tracks, the post argues that science routinely proceeds from observable phenomena (missing momentum, distorted jet shapes, anomalous events) to unobservable reality.

Takeaways for Partially Observable Systems

1. Are you fully exploiting the information in the observable part? 2. Do you have a theoretical model constraining the space of the unobservable? 3. Does your inference method quantify uncertainty? 4. Have you considered alternative hypothesis scenarios?

Even facing an "invisible" reality, careful experiment design, strong theoretical frameworks, and modern machine learning allow us to reconstruct an unseen picture of the universe from visible clues.