Domain-Informed Neural Networks for Interaction Localization Within Astroparticle Experiments

Area of Science:

• Experimental particle physics
• Dark matter research
• Machine learning applications

Background:

• Particle interaction reconstruction in time-projection chambers (TPCs) is crucial for experiments.
• Multilayer perceptrons (MLPs) are used for TPC reconstruction but lack domain knowledge.
• Existing black-box models do not leverage prior scientific understanding of detector physics.

Purpose of the Study:

• To develop a novel neural network architecture for particle physics.
• To incorporate prior knowledge of detector physics and signal characteristics into neural network design.
• To improve the efficiency and interpretability of particle interaction localization in TPCs.

Main Methods:

• Proposed a Domain-informed Neural Network (DiNN) architecture.
• Encoded detector geometry and signal properties into the neural network's feature encoding and output layers.
• Limited neuron receptive fields in initial layers and modified output layers with geometric transformations.

Main Results:

• The DiNN architecture significantly reduces model parameters by 60% compared to MLPs.
• Achieved comparable localization performance to existing MLP-based methods.
• Demonstrated a novel approach for integrating domain knowledge into deep learning models.

Conclusions:

• Domain-informed neural networks offer a more efficient and potentially more performant alternative to black-box models in particle physics.
• The DiNN architecture provides a framework for incorporating specific scientific knowledge into machine learning models.
• This approach paves the way for future developments in AI for scientific discovery, particularly in dark matter research.