[Paper Review] Nanosecond anomaly detection with decision trees and real-time application to exotic Higgs decays

Stephen Roche, Quincy Bayer+5 more|arXiv (Cornell University)|Apr 7, 2023

Particle physics theoretical and experimental studies103 references4 citations

TL;DR

This paper presents a real-time, FPGA-based anomaly detection system using a forest of deep decision trees to implement an autoencoder for high-energy physics, achieving 25 ns inference latency—matching the LHC's collision rate—while using only percent-level FPGA resources. The method detects exotic Higgs decays that evade conventional threshold-based triggers, enabling discovery of new physics in real time with minimal latency and resource usage.

ABSTRACT

We present an interpretable implementation of the autoencoding algorithm, used as an anomaly detector, built with a forest of deep decision trees on FPGA, field programmable gate arrays. Scenarios at the Large Hadron Collider at CERN are considered, for which the autoencoder is trained using known physical processes of the Standard Model. The design is then deployed in real-time trigger systems for anomaly detection of unknown physical processes, such as the detection of rare exotic decays of the Higgs boson. The inference is made with a latency value of 30 ns at percent-level resource usage using the Xilinx Virtex UltraScale+ VU9P FPGA. Our method offers anomaly detection at low latency values for edge AI users with resource constraints.

Motivation & Objective

• To develop a real-time anomaly detection system for high-energy physics that operates within the 25 ns collision window of the LHC.
• To overcome the latency and resource constraints of existing neural network-based anomaly detectors in FPGA-trigger systems.
• To enable detection of previously undetected exotic Higgs decays that fail conventional threshold-based triggers.
• To demonstrate robustness of the method under signal-contaminated training scenarios typical of real-world LHC data.

Proposed method

• The method employs a tree-based autoencoder architecture using deep decision trees (DDT) with parallel decision paths (PDP) for ultrafast inference on FPGA.
• The encoder maps input data to a latent space via threshold-based splits in a deep decision tree, while the decoder reconstructs the input from the latent representation.
• The system is trained on Standard Model processes to learn normal event patterns, enabling detection of deviations as anomalies.
• Firmware is implemented using the fwXmachina framework, enabling efficient, low-latency deployment on FPGAs with minimal DSP usage.
• The architecture uses a star-coder (⋆-coder) design that enables simultaneous encoding and decoding through parallel decision paths, avoiding explicit latent layer storage.
• The method is validated using simulated LHC datasets, including a benchmark for exotic Higgs decays and a broader LHC physics dataset with multiple BSM models.

Experimental results

Research questions

• RQ1 Can a decision tree-based autoencoder achieve real-time anomaly detection at 25 ns latency on FPGA for LHC trigger systems?
• RQ2 How effective is the method in detecting exotic Higgs decays that fail conventional single-lepton trigger thresholds?
• RQ3 How does the system perform when trained on data contaminated with signal events, simulating real-world deployment conditions?
• RQ4 What is the resource efficiency of the method in terms of FPGA utilization and power consumption?

Key findings

• The system achieves inference latency of 25 ns, matching the LHC's 40 MHz collision rate, making it the lowest-latency anomaly detection method reported for edge AI in high-energy physics.
• The method uses only percent-level FPGA resources, enabling deployment in tight resource-constrained trigger systems.
• The autoencoder successfully detects exotic Higgs decays to two BSM pseudoscalars decaying to leptons, events that are discarded by conventional threshold-based triggers.
• The anomaly detection performance remains robust even when training data contains up to 10% signal contamination, demonstrating practical viability.
• The method achieves a data compression-decompression factor of ~300 in a test case using MNIST, indicating strong reconstruction fidelity for complex input spaces.
• The framework is generalizable and applicable to other edge AI applications requiring ultra-low latency and minimal resource usage.