| ISSUE 09/30 · COMPUTER SCIENCE |
~4 MIN |
NOW / CERN + MACHINE LEARNING · WEEK 02 — MACHINES UNDER PRESSURE
How software rebuilds a particle collision
START HERE
When tiny pieces of matter collide at CERN, they do not leave a normal photograph. Sensors record separate electronic signals at known positions and deposits of energy inside a layered machine. Reconstruction is the software process of working backward from those raw signals to estimate which particles passed through and how they moved. If the software joins the wrong marks together, it can invent a particle, lose one, or count the same energy twice.
|
/ WHY NOW
When two protons—small particles found inside atomic nuclei—collide at CERN, they create a spray of short-lived particles. Sensors record raw hits, energy and timing; software must fit some hits into tracks before it can name a particle. Machine learning is software that learns patterns from examples instead of receiving every rule by hand. In February 2026, CMS, one of CERN’s large detector experiments, reported a learned system that rebuilt full collisions faster and, in tests, more precisely than its hand-built predecessor.
|
/ THE IDEA
Reconstruction works backward: the detector shows effects and software infers causes. A charged particle may leave a track in one layer and an energy cluster in another. The program decides which traces belong together and what particle explains them. Particle flow is the broader job of combining several detector systems to identify particles and estimate their momentum—the amount and direction of their motion. Traditional versions use chains of expert rules. The reported learned model starts from reconstructed tracks and calorimeter clusters, then maps that collection to particle candidates; raw detector hits are processed upstream.
THE FORMAL IDEA
estimated particle set = model(tracks + energy clusters)
| input = reconstructed tracks plus clusters of energy recorded in calorimeters | | output = several particle candidates rather than one label | | each candidate includes a likely identity and estimated motion |
|
RUN THE TINY EXAMPLE
One track, one flash
Input A: a track points to an energy cluster → link them as one charged-particle candidate Input B: move that cluster away from the track → do not link those two traces Output B may contain a charged-track candidate plus a separate neutral-energy candidate
|
If the linking is wrong, energy can be counted twice or assigned to the wrong particle. Better association improves every later physics measurement.
/ SO WHAT?
The lesson reaches beyond CERN. Medical scans, telescopes and self-driving sensors also ask software to infer hidden objects from incomplete measurements. The visible data is not the thing itself; it is evidence filtered through an instrument.
ONE CAVEAT |
| A model trained on simulation can inherit gaps between the simulation and the real detector. Speed or benchmark accuracy does not remove the need for calibration, uncertainty estimates and tests on real data. |
KEEP THIS
Particle reconstruction is the disciplined act of turning detector effects back into their most likely causes.
|
NEXT: Finding a particle from what it left behind
|