Federated Learning with Quantum Enhanced LSTM for Applications in High Energy Physics

1. Problem Statement

High Energy Physics (HEP) experiments, such as those at the Large Hadron Collider (LHC), generate petabyte-scale datasets annually. Extracting rare physical events (e.g., Supersymmetry or SUSY signals) from overwhelming background noise requires highly complex machine learning models. However, current approaches face three critical challenges:

• Resource Constraints: Training large-scale deep learning models on supercomputers incurs massive energy costs and computational overhead.
• Data Privacy & Distribution: HEP data is distributed across global institutions (colliders and detectors). Centralizing this data for training raises privacy concerns and logistical hurdles.
• Quantum Limitations: While Quantum Machine Learning (QML) offers superior representational capabilities for high-dimensional data, current Noisy Intermediate-Scale Quantum (NISQ) devices are limited by noise, qubit count, and error correction, making them unsuitable for training on massive datasets independently.

The paper proposes a solution that combines Federated Learning (FL) to distribute the workload and Quantum-Enhanced Long Short-Term Memory (QLSTM) to maximize learning efficiency with minimal data and parameters.

2. Methodology

The authors propose a Hybrid Quantum-Classical Federated Learning Framework tailored for HEP applications.

A. The Quantum-Enhanced LSTM (QLSTM) Model

Instead of using a standalone Variational Quantum Circuit (VQC), the authors design a hybrid architecture where quantum circuits replace specific components within a classical LSTM cell.

• Architecture:
  1. Input Projection: Classical input vectors ($X_t$) are projected into a quantum embedding space using a linear layer ($z_t = W_c X_t + b_c$).
  2. Quantum Encoding: Features are encoded into quantum states using Angle Encoding (mapping features to rotation angles of single-qubit gates).
  3. Variational Quantum Circuit (VQC): The encoded state passes through a parametrized circuit with entangling gates (e.g., CNOT) and trainable rotation gates ($R_y, R_z$). This layer learns complex, non-linear correlations in the feature space.
  4. Measurement & Reconstruction: The quantum state is measured in the Z-basis, and the results are mapped back to classical space via a linear layer.
  5. LSTM Integration: These quantum-enhanced features drive the standard LSTM gates (forget, input, output, and cell gates) to update memory and hidden states.
• Advantage: This design leverages the quantum model's ability to map high-dimensional feature spaces efficiently while using the LSTM's temporal correlation learning capabilities.

B. Federated Learning (FL) Setup

To address data distribution and privacy:

• Nodes: Different HEP detectors or institutions act as "clients," each holding local data and training a local QLSTM model.
• Server: A central server aggregates the model weights (parameters) from all nodes to create a global model, which is then redistributed.
• Workflow: Local training occurs on distributed nodes; only model updates (weights) are shared, preserving data privacy and reducing bandwidth requirements.

C. Data Encoding Strategy

The paper utilizes Angle Encoding rather than Amplitude Encoding. While Amplitude Encoding offers better compression ($O(\log N)$), empirical studies cited suggest Angle Encoding often yields higher classification accuracy on specific datasets, making it more suitable for the current NISQ hardware constraints.

3. Key Contributions

1. Novel Hybrid Architecture: The design of a QLSTM specifically for HEP, integrating VQCs into recurrent neural networks to learn complex correlations with fewer parameters.
2. Federated Quantum Framework: The proposal of a QFL (Quantum Federated Learning) system for HEP, enabling collaborative learning across global institutions without sharing raw data.
3. Efficiency Breakthrough: The demonstration that a hybrid quantum-classical model can achieve state-of-the-art performance with drastically reduced resource requirements (data and parameters) compared to classical deep learning baselines.
4. Empirical Validation: Extensive experiments on the SUSY dataset (LHCb) comparing QLSTM against standalone VQC, classical LSTM, and existing literature.

4. Experimental Results

The experiments were conducted using the SUSY dataset (5 million rows, though only 20k were used for simulation due to hardware limits) with two feature sets: 18 features (full) and 7 features (selected significant ones).

• Performance Metrics: The models were evaluated using AUC (Area Under the Curve) and Test Accuracy.
• Comparison with Baselines:
  • QLSTM vs. VQC: QLSTM significantly outperformed standalone VQC models.
    • QLSTM (18 features): AUC $\approx$ 0.880.
    • VQC (18 features): AUC $\approx$ 0.823.
  • QLSTM vs. Classical LSTM: QLSTM outperformed classical LSTM, showing that the quantum layer adds value in capturing non-linear correlations.
    • QLSTM: AUC $\approx$ 0.880.
    • LSTM: AUC $\approx$ 0.862.
  • Comparison to State-of-the-Art: The QLSTM results were comparable (within $\pm 1\%$) to the best classical deep learning benchmarks reported in previous literature (e.g., Baldi et al.), despite using a much smaller model.
• Federated Learning Impact:
  • Performance degradation in the federated setting (3 nodes) was minimal ($\Delta < 1\%$).
  • QLSTM showed greater robustness to data splitting than simple VQC models, maintaining high AUC even with distributed data.
• Resource Efficiency (The "100x" Improvement):
  • Parameters: The proposed model has < 300 parameters, compared to ~300,000 in comparable classical deep learning models.
  • Data: The model achieved comparable performance using only 20,000 data points, whereas baseline models typically require millions (e.g., 5M).
  • Conclusion: This represents a 100x improvement in data and parameter efficiency.

5. Significance and Future Work

Significance:

• Scalability: The framework proves that quantum-enhanced models are viable for large-scale scientific applications even on current NISQ hardware, provided they are integrated into hybrid architectures and distributed via FL.
• Cost Reduction: By requiring significantly less data and fewer parameters, the approach reduces the computational cost and energy consumption associated with training HEP models.
• Privacy: The FL component enables global collaboration on sensitive physics data without compromising data sovereignty.

Future Work:

• Heterogeneous Data: Extending the framework to handle non-IID (non-independent and identically distributed) data across different detectors.
• Hardware Deployment: Moving from simulation (PennyLane) to actual quantum hardware to study the impact of real-world noise characteristics.
• Encoding Optimization: Exploring other compact encoding techniques to further enhance QML performance in particle physics.

In summary, the paper successfully demonstrates that a Federated QLSTM is a highly efficient, privacy-preserving, and high-performing solution for High Energy Physics classification tasks, offering a path forward for integrating quantum computing into large-scale scientific workflows.