Abstract
Antihydrogen is at the forefront of antimatter research at the CERN Antiproton Decelerator. Experiments aiming to test the fundamental CPT symmetry and antigravity effects require the efficient detection of antihydrogen annihilation events, which is performed using highly granular tracking detectors installed around an antimatter trap. Improving the efficiency of the antihydrogen annihilation detection plays a central role in the final sensitivity of the experiments. We propose deep learning as a novel technique to analyze antihydrogen annihilation data, and compare its performance with a traditional track and vertex reconstruction method. We report that the deep learning approach yields significant improvement, tripling event coverage while simultaneously improving performance by over 5% in terms of Area Under Curve (AUC).
Introduction
In antimatter physics, detecting the annihilation of an antihydrogen atom is crucial for testing fundamental symmetries like Charge, Parity, and Time reversal (CPT) and the gravitational behavior of antimatter. Upon annihilation, antihydrogen produces several charged pions. Traditionally, these are detected via "track and vertex reconstruction," where hits in a detector are fitted into tracks to find their intersection point (the vertex).
However, the experimental environment is challenging. Antihydrogen traps contain vacuum walls, electrodes, and thermal isolation layers that cause "multiple Coulomb scattering," smearing pion paths. In the ASACUSA experiment, traditional vertex-finding (VF) algorithms often fail due to this noise, failing to reconstruct a vertex for approximately 75% of events and resulting in significant data loss.
Methods
To overcome these limitations, we propose an end-to-end deep learning approach that classifies events directly from low-level detector data, bypassing explicit vertex reconstruction.
Architecture
The proposed Deep Convolutional Neural Network (DNN) utilizes two parallel 1D-convolutional pathways. One pathway processes data from the azimuthal ($\phi$) dimension, while the other handles the axial ($Z$) dimension. Each pathway consists of five 1D convolutional layers followed by max-pooling to achieve translational invariance. These features are merged into dense layers for final classification.
Data Augmentation and Training
To ensure robustness against the unknown axial distribution of antihydrogen in real traps, we employed "translation augmentation." Simulated events were randomly shifted along the $Z$-axis during training, forcing the network to learn the characteristic shape of annihilation events rather than specific locations.
Dataset
The study utilized a dataset of two million events (1 million signal, 1 million background) generated using the Geant4 toolkit. This simulation incorporated the full material budget of the ASACUSA cryogenic vacuum trap and the Asacusa Micromegas Tracker (AMT). The dataset is publicly available for further research.
Results
The DNN approach demonstrated significant improvements over traditional methods across multiple metrics.
Superior Classification Performance
On the subset of events where the traditional VF algorithm successfully found a vertex, the DNN still outperformed it, achieving an AUC of 0.872 compared to the VF's 0.762. This indicates that the DNN captures subtle correlations in raw hit patterns that traditional geometric models miss.
Massive Efficiency Gain
The most impactful result is the increase in event coverage. Because the DNN does not require a reconstructed vertex, it can provide predictions for 100% of events. In practice, it effectively tripled the event coverage, reaching 73% of all events (compared to 25% for VF) while maintaining high performance (82% AUC on the union of all detectable events).
Discussion and Conclusion
Deep learning offers a powerful alternative to traditional reconstruction in high-energy physics. By processing raw detector data, the DNN recovers information lost during traditional track fitting and is more robust to Coulomb scattering noise.
While these results are promising, future work must address the "sim-to-real" gap by validating the model on experimental data. Additionally, while the approach is general, the specific architecture is currently optimized for the ASACUSA trap geometry. Further research into model interpretability will also be valuable for understanding the underlying particle dynamics identified by the network.
Summary updated