Abstract: Under extreme operating conditions, characterized by high particle multiplicity and heavily overlapping shower energy deposits, classical particle flow algorithms encounter pronounced limitations in resolution, efficiency, and accuracy. To address this challenge, this paper proposes and systematically evaluates a deep learning reconstruction framework: For multichannel sparse features, we design a hybrid loss function combining weighted mean squared error with structural similarity index, effectively balancing pixel-level accuracy and structural fidelity. By integrating 3D convolutions, Squeeze-and-Excitation channel attention, and Offset self-attention modules into baseline convolutional neural networks, we enhance the model's capability to capture cross-modal spatiotemporal correlations and energy-displacement nonlinearities. Validated on custom-constructed simulation data and Pythia jet datasets, the framework's 90K-parameter lightweight variant approaches the performance of 5M-parameter baselines, while the 25M-parameter 3D model achieves state-of-the-art results in both interpolation and extrapolation tasks. Comprehensive experiments quantitatively evaluate component contributions and provide performance-parameter trade-off guidelines. All core code and data processing scripts are open-sourced on a GitHub repository to facilitate community reproducibility and extension.
1. Introduction
In high-energy physics experiments, such as those conducted at the Large Hadron Collider (LHC), Particle Flow (PFlow) algorithms are critical for reconstructing particle trajectories and energies from detector data. However, as collision intensity increases, detectors face high particle multiplicity and overlapping shower energy deposits. Classical PFlow algorithms often struggle in these dense environments, leading to degraded resolution and accuracy. While deep learning offers a potential solution, many existing models are computationally prohibitive or fail to account for the non-linear energy-displacement relationships caused by magnetic fields.
2. Methodology
The proposed framework utilizes a multi-component architecture designed to capture complex spatial correlations and enhance feature representation.
2.1 Architectural Components
The core of the framework is a multi-layer convolutional encoder-decoder structure. Key innovations include:
- 3D Convolutional Neural Networks (3D CNNs): These treat detector layers as depth, preserving spatial correlations across different detector components (e.g., EMCal and HCal).
- Squeeze-and-Excitation (SE) Blocks: These provide a channel attention mechanism, allowing the network to prioritize informative signals and suppress noise.
- Offset Self-Attention: This module establishes global mappings between energy and displacement, accounting for the curved paths of particles in magnetic fields.
2.2 Hybrid MIX Loss Function
To address the sparsity of detector data, a hybrid "MIX" loss function was developed. It combines Weighted Mean Squared Error (WMSE) with the Structural Similarity Index (SSIM). While WMSE ensures pixel-level accuracy, SSIM preserves the structural "shape" of particle showers, which is often lost when using standard MSE alone.
3. Experiments and Results
The framework was evaluated using custom simulation data and Pythia 8 jet datasets to test both interpolation and extrapolation capabilities.
3.1 Model Efficiency and Performance
A lightweight variant of the model, the SEBlockCNN with only 90K parameters, demonstrated performance comparable to a 5M-parameter baseline CNN. This represents a significant reduction in computational overhead, making it suitable for real-time physics triggers. For maximum precision, a 25M-parameter 3D model achieved state-of-the-art results, particularly in handling high-energy jets.
3.2 Impact of the MIX Loss
Quantitative analysis revealed that the MIX loss function significantly improves spatial localization. As shown in Figure 6, the MIX loss minimizes the distance difference (barycenter shift) compared to standard MSE or WMSE alone, leading to more accurate energy reconstruction.
4. Conclusion
This research presents a lightweight yet powerful deep learning framework for particle flow reconstruction. By integrating attention mechanisms and a structural-aware loss function, the model achieves high accuracy with minimal parameters. Future work will focus on adaptive loss weighting and real-time implementation on hardware-level triggers like FPGAs.
Summary updated