Sub-Microsecond Machine Learning for Real-Time Physics Triggers

This direction is split into other directions: triggerless data acquisition, high-throughput data filtering. Its last version stays readable and citable here.

Version 1 of 1140 papers161 conceptsfrozen 2026-10-04

In high-energy physics, the challenge is not just finding a needle in a haystack—it is finding the needle while the haystack is being thrown at you at 40 million frames per second. Experiments like those at the Large Hadron Collider (LHC) generate data at rates far exceeding the capacity of any storage system. To survive this data deluge, researchers must make irreversible "keep or discard" decisions in real-time.

This filtering process, known as the trigger, is shifting from simple, hard-coded logic to sophisticated machine learning models that operate within nanoseconds.

The Latency Wall

The fundamental constraint in experiments like LHCb or Belle II is the "latency budget." When particles collide, the detector's buffers can only hold the resulting data for a few microseconds before they are overwritten by the next event.

Traditional software running on general-purpose CPUs is too slow to process these events in time. As Javier Duarte and Jennifer Ngadiuba highlight in their work on Efficient AI, this temporal constraint requires a move toward hardware-accelerated, "leaner" AI models. If a model takes even a millisecond to run, the experiment loses thousands of potentially Nobel-winning events.

Translating Neural Networks to Silicon

To meet these microsecond requirements, researchers are moving inference directly onto Field-Programmable Gate Arrays (FPGAs) and Application-Specific Integrated Circuits (ASICs). The core innovation enabling this is the hls4ml library, developed by researchers including Farah Fahim and Vladimir Lončar.

hls4ml acts as a bridge, automatically translating high-level machine learning models (like those in PyTorch or Keras) into hardware description languages. It employs three main strategies to achieve sub-microsecond speeds:

  1. Quantization: Reducing the precision of numbers (e.g., from 32-bit floats to 6-bit integers) to save space and power.
  2. Pruning: Removing unnecessary connections in a neural network that don't contribute to the final decision.
  3. Parallelism: Exploiting the FPGA's architecture to process different parts of the neural network simultaneously.

From Pixels to Graphs and Equations

The choice of model architecture is critical for balancing accuracy and speed. While standard neural networks treat detector data like images, newer approaches treat them as relational data:

  • Graph Neural Networks (GNNs): Particle tracks are essentially "dots" (hits) that need to be connected. Gage DeZoort and Javier Duarte have shown that GNNs can reconstruct these paths by treating hits as nodes and potential tracks as edges. To make this work in real-time, architectures like LL-GNN optimize matrix multiplications to fit within the strict latency windows of the LHC.
  • Symbolic Regression (SR): Perhaps the most extreme form of optimization is replacing a neural network entirely with a simple algebraic equation. Adrian Alan Pol and Ho Fung Tsoi demonstrated that Symbolic Regression on FPGAs can reduce latency to a staggering 5 nanoseconds. By finding the simplest mathematical expression that fits the data, they achieve 90% of a complex model's accuracy with a fraction of the computational cost.
  • Transformers: Even complex architectures like Transformers are being adapted. By optimizing Multi-Head Attention layers for hardware, Zhixing Jiang and colleagues achieved Transformer inference in under 2 microseconds, enabling these models to be used in gravitational wave detection and particle triggers.

The Shift to Streaming Readout

As luminosity (the rate of collisions) increases, even the best hardware triggers face a bottleneck. This has led to the concept of Streaming Readout or "trigger-less" systems. Instead of a hardware signal telling the detector when to "snap a photo," the detector streams data continuously.

In this paradigm, described in the context of the DarkLight experiment and LHCb Run 3, the filtering happens in a massive farm of GPUs or FPGAs that perform real-time reconstruction. C. Prouvé and C. Hasse demonstrated that using DeepSet architectures for inclusive flavor tagging can provide a 5x speedup, making software-only triggers viable for high-rate environments.

Why It Works: Hardware-Algorithm Co-Design

The success of this research direction stems from co-design: the simultaneous optimization of the physics algorithm and the underlying hardware. By understanding the physical symmetries of the data (as seen in symmetry-preserving attention networks by Haoyang Li) and the specific constraints of FPGAs, researchers can build filters that are both "open-minded" enough to catch new physics and fast enough to never miss a beat.

Papers in this version

Showing 60 of 140 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_sub_microsecond_machine_learning_for_rea_97a9f9c3,
  title = {Sub-Microsecond Machine Learning for Real-Time Physics Triggers},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sub-microsecond-machine-learning-for-real-time-physics-triggers/rdr_97a9f9c3ff6941daa088781c73bcc1d7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_sub_microsecond_machine_learning_for_rea_97a9f9c3_v1,
  title = {Sub-Microsecond Machine Learning for Real-Time Physics Triggers},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sub-microsecond-machine-learning-for-real-time-physics-triggers/rdr_97a9f9c3ff6941daa088781c73bcc1d7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}