Classical Hardware Acceleration of Quantum Autoencoders for Real-Time Anomaly Detection in Collider Experiments

Technical Summary: Classical Hardware Acceleration of Quantum Autoencoders for Real-Time Anomaly Detection in Collider Experiments

Problem Statement
High Energy Physics (HEP) experiments, such as the Large Hadron Collider (LHC), face extreme data throughput challenges, generating tens of terabytes per second. To manage this, multi-layered trigger systems must filter events in real-time, reducing data rates to levels suitable for storage. While classical machine learning (ML) is ubiquitous in this domain, Quantum Machine Learning (QML) offers a promising alternative for detecting rare Beyond the Standard Model (BSM) signals. These signals often manifest as subtle, high-dimensional, long-range correlations that QML algorithms can represent efficiently with fewer parameters. However, deploying QML in real-time collider triggers requires emulating quantum circuits on low-latency hardware accelerators, specifically Field-Programmable Gate Arrays (FPGAs), which are standard in trigger systems (e.g., ATLAS Level-1). The challenge lies in demonstrating that QML models can achieve physics performance comparable to state-of-the-art classical methods while satisfying the strict resource and timing constraints of FPGA deployment.

Methodology
The study evaluates Variational Quantum Autoencoder (QAE) models for anomaly detection (AD) using a dataset of simulated proton-proton collisions at the LHC. The dataset includes a background of Quantum Chromodynamics (QCD) multijet events and four distinct BSM signal processes (involving scalar bosons and leptoquarks).

1. Particle Embedding: Events are represented by 56 real-valued features describing 19 reconstructed objects (electrons, muons, jets, and missing transverse energy). Following a "one-particle-one-qubit" convention, these features are encoded into quantum states using parameterized single-qubit rotation gates ($R_X, R_Y, R_Z$). To accommodate FPGA resource limits, the 19-qubit register is decomposed into smaller sub-blocks: 1 qubit for missing energy, 4 for electrons, 4 for muons, and two independent 5-qubit blocks for the ten jets. This decomposition avoids the exponential memory cost of a fully entangled 10-jet state vector.
2. Model Architectures:
   • Hybrid Quantum-Classical VAE: This model replaces the first dense layer of a classical Variational Autoencoder with block-structured Parameterized Quantum Circuits (PQCs). The quantum blocks encode particle features and apply entanglement, outputting a 32-dimensional feature vector of measured Pauli observables. This vector is processed by classical dense layers to map to a 3-dimensional Gaussian latent space. The anomaly score is the clipped Kullback-Leibler (KL) divergence.
   • Fully Quantum Autoencoder (QAE): This model removes all classical layers. It uses the same particle-type block partitioning but processes data entirely within quantum circuits. The objective is to compress background events into latent qubits while pushing "trash" wires toward the $|0\rangle$ state. Anomaly scores are derived from the excitation probabilities of these trash wires, normalized against background statistics.
3. Training: Both models are trained on background-only data using the Adam optimizer. The Hybrid VAE minimizes a standard ELBO loss, while the QAE minimizes a "z-loss" to suppress trash wire excitations on background events.
4. FPGA Synthesis: Trained models are quantized to fixed-point arithmetic and synthesized for the Xilinx Alveo U200 FPGA. The synthesis targets a 5.5 ns clock cycle. The authors utilized agentic AI tools to optimize resource usage, ensuring the design fits within a single FPGA Super Logic Region (SLR) while minimizing latency.

Key Contributions

• First FPGA Implementation for HEP Triggers: This work presents one of the first implementations of QML models specifically designed for HEP trigger applications on FPGAs.
• Hybrid and Fully Quantum Architectures: The study compares a hybrid quantum-classical VAE against a fully quantum autoencoder, demonstrating that both can be effectively emulated on classical hardware.
• Resource-Constrained Optimization: The paper details the decomposition of large quantum registers into smaller sub-blocks to make simulation and FPGA synthesis feasible within real-time constraints.
• End-to-End Benchmarking: The study provides a comprehensive evaluation of physics performance (AUC and TPR) alongside hardware metrics (latency, LUTs, DSPs, FFs).

Results

• Physics Performance: Both models achieve anomaly detection performance comparable to or exceeding state-of-the-art classical ML methods. The Hybrid VAE generally shows a higher Area Under the Curve (AUC), while the QAE is specifically optimized for the low false-positive rate (FPR) regime relevant to triggers. Both models maintain a competitive True Positive Rate (TPR) at a fixed background efficiency of $10^{-5}$ across all four BSM signal models.
• Quantization and Synthesis: Fixed-point quantization analysis reveals that while AUC can be artificially inflated or stable at lower precisions, the TPR@$10^{-5}$ metric is a more faithful indicator of deployment performance.
• Latency and Resources:
  • The Hybrid VAE achieves a latency of 6.1 µs and utilizes 208,976 LUTs and 1,528 DSPs.
  • The QAE achieves a significantly lower latency of 0.47 µs (under 1 µs) with 249,777 LUTs and 1,234 DSPs.
  • Both models fit within a single FPGA SLR.
• Hardware Comparison: In per-event inference latency, the FPGA implementation is the fastest option for a batch size of one (the trigger operating regime). However, for larger batch sizes, multi-threaded CPU and GPU implementations surpass the FPGA in throughput, consistent with the observation that FPGA emulation is advantageous for small quantum registers where per-event work is limited.

Significance and Claims
The paper claims that this work demonstrates the feasibility of deploying QML models for real-time anomaly detection in future collider experiments without sacrificing sensitivity to rare BSM signatures. By successfully synthesizing these models onto FPGAs and meeting strict latency and resource budgets, the study suggests that quantum-ready models can be integrated into today's classical data acquisition pipelines. The authors position this as a step toward "quantum readiness" for collider infrastructure, enabling higher-capability models in trigger systems while leveraging the efficiency of quantum algorithms represented on classical hardware accelerators. The work does not claim to replace classical methods but rather to complement them, offering a path to fundamentally new BSM phase space exploration under the constraints of real-time data filtering.