Five hundred thousand times per second, the Electron-Ion Collider (EIC) will record a collision. At that rate, machine learning will sort, filter, and reconstruct what’s happening inside the detector. That requirement shaped the entire facility’s design.

The collider being built at Brookhaven National Laboratory in New York is the first of its kind – the first particle collider with AI and machine learning integrated into both its accelerator and detector systems from the start. It is a joint project between Brookhaven and the Department of Energy’s Thomas Jefferson National Accelerator Facility, with more than 300 institutions collaborating worldwide.

The price tag sits between $1.7 billion and $2.8 billion. Operations are targeted for the mid-2030s.

Teaching the accelerator to tune itself

Earlier particle physics facilities, including Brookhaven’s own Relativistic Heavy Ion Collider, which shut down in February 2026, incorporated AI tools years after construction. For the EIC, a multi-institutional group called EIC-BeamAI is developing machine learning systems using live accelerator hardware at Brookhaven.

The challenge is significant: keeping a particle accelerator stable means managing tens of thousands of parameters simultaneously, across two beams running in opposite directions around a 2.4-mile ring at close to the speed of light.

“It’s very difficult for a human being to keep on top of all these settings and beam characteristics all the time,” said Georg Hoffstaetter de Torquat, a Cornell University professor with a joint appointment at Brookhaven. “With machine learning, what we write is essentially computer supervision — the system monitors conditions and adjusts controls automatically.”

BeamAI has already proven the concept. In RHIC’s pre-accelerators, machine learning algorithms matched the beam quality that expert human operators typically achieve.

The system also produces a digital twin of the accelerator, a real-time virtual model that lets researchers test changes without touching the live machine. That same twin can catch abnormal magnet behavior early enough to trigger a controlled shutdown before anything is damaged.

Rethinking detector design

Building a particle detector means running an enormous number of simulations before a single component is manufactured: testing geometry, materials, and configuration against countless collision scenarios. A DOE-supported project called AID2E, spanning Brookhaven, The Catholic University of America, Duke University, Jefferson Lab, and William & Mary, is applying machine learning to that process.

Algorithms trained to predict how design changes affect particle identification let researchers move through far more configurations than standard simulation workflows permit, at lower computing cost and energy use.

The data problem

When the EIC goes online, its detector — a house-sized instrument called ePIC — will produce up to 100 gigabits of data per second. AI-driven systems will sort that stream in real time, separating signal from noise as collisions occur. Deep learning models will then reconstruct what happened in each event: translating the faint traces particles leave inside the detector into usable measurements of energy and momentum.

A related Brookhaven project, published in the journal Patterns, demonstrated an algorithm capable of compressing collision data at scale without losing the granularity required by physics analysis — built and tested on RHIC hardware.