André Anjos

High-Energy Physics — ATLAS Trigger (CERN)

Earlier work at CERN's ATLAS experiment on real-time trigger and data-acquisition systems, and neural-network methods for particle recognition.

Partnerships

CERN, ATLAS ExperimentFederal University of Rio de Janeiro

Earlier-career work at CERN (2001 to 2010).

I spent close to a decade at CERN’s ATLAS experiment, working on the Trigger and Data Acquisition systems that decide, in real time, which of the Large Hadron Collider’s collisions are worth recording. That infrastructure helped make discoveries like the Higgs boson possible. In parallel, I explored neural-network and statistical methods for recognising particles as the data streamed in.

Major achievements

I contributed to the design, commissioning, and operation of the ATLAS High-Level Trigger, the software stage that decides which collisions are kept. The scale sets the problem: the collider delivers collisions at 40 MHz and the detector produces around 1.5 megabytes per event, some 60 terabytes per second of raw data, of which only a small fraction can ever be written to storage. The architecture we helped build and validate splits that decision across two software stages running on processor farms, and keeps it affordable by having the hardware trigger point at regions of interest, so the second stage reads under 2% of an event instead of all of it. My work concerned the dataflow and supervision of that system — how events move through the farms, and how the farms themselves are configured, controlled, and monitored — and later a software environment that automated configuring and running the trigger and dataflow farms, so that a system of this size could be tested and redeployed by people ranging from casual testers to final deployers. Some of those components stayed in use well after I left.

In parallel, my doctoral work applied neural networks and statistical methods to separating electrons from jets in the online filter, from calorimeter data alone and within a hard real-time budget. The obstacle was dimensionality: a calorimeter region of interest carries far too many channels to feed a network directly at these rates. Organising the deposited energy into concentric ring sums around the shower axis compressed the input while preserving the structure that distinguishes an electron from a hadronic jet, and so improved discrimination rather than merely making it cheaper — the resulting network reached 97% electron efficiency at a 3% false-alarm rate, with the full discrimination chain executing in under 500 microseconds. More than anything, this period taught me how large scientific software is really built, and how much of a physics result rests on infrastructure nobody outside the collaboration ever sees. That is what drew me toward reproducible, well-engineered computing in the years that followed.

Key publications