Reproduce “Domain-informed neural networks for interaction localization within astroparticle experiments” on OpenResearch

About this paper

Shixiao Liang, Aaron Higuera, Christina Peters, Venkat Roy, Waheed U. Bajwa, Hagit Shatkay, Christopher D. Tunnell

Domain-informed Neural Networks (DiNNs) precisely localize particle interactions (S2 signals) in astroparticle time-projection chambers by integrating detector geometry and signal characteristics directly into the network architecture. This approach yields comparable or improved localization accuracy relative to standard MLPs, while reducing trainable parameters by 60% and guaranteeing physically valid output predictions.

Key insights

• Integrating domain knowledge, such as signal locality and detector geometry, directly into a neural network's architecture can drastically improve efficiency and physical consistency.
• Graph-constrained hidden layers enable sparse connectivity, significantly reducing trainable parameters and leveraging inherent spatial correlations in data from irregularly arranged sensors.
• A geometry-constrained output layer ensures that all localization predictions are physically valid, preventing the reporting of impossible interaction coordinates outside the detector.

Summary by alphaXiv.