Photon detection probability prediction using one-dimensional generative neural network

TL;DR

Large-scale liquid argon detectors require detailed photon transport simulations, which are computationally expensive with Geant4. The authors introduce a one-dimensional generative neural network (Genn) that conditions on the scintillation vertex and uses an OuterProduct-layer to produce a 1D hit-pattern vector for photon detectors, trained with a variational KL loss (D_vKL) on Geant4 data. The model achieves 20–50× faster CPU inference while maintaining Geant4-level accuracy in both ProtoDUNE-like and DUNE-like geometries, with favorable memory usage. The approach offers a scalable, CPU-friendly alternative for rapid photon-detection probability predictions and could generalize to other 1D-vector–based simulation tasks in large LAr detectors.

Abstract

Photon detection is important for liquid argon detectors for direct dark matter searches or neutrino property measurements. Precise simulation of photon transport is widely used to understand the probability of photon detection in liquid argon detectors. Traditional photon transport simulation, which tracks every photon using theGeant4simulation toolkit, is a major computational challenge for kilo-tonne-scale liquid argon detectors and GeV-level energy depositions. In this work, we propose a one-dimensional generative model which efficiently generates features using an OuterProduct-layer. This model bypasses photon transport simulation and predicts the number of photons detected by particular photon detectors at the same level of detail as theGeant4simulation. The application to simulating photon detection systems in kilo-tonne-scale liquid argon detectors demonstrates this novel generative model is able to reproduceGeant4simulation with good accuracy and 20 to 50 times faster. This generative model can be used to quickly predict photon detection probability in huge liquid argon detectors like ProtoDUNE or DUNE.

Figures (6)

• Figure 1: General Genn architecture. ${pos_y}$-layer and ${pos_z}$-layer are used to predict the photon distribution on the photon detection system, while ${pos_x}$-layer is a normalization factor for the number of total detected photons. $\mathrm{OuterProduct}$ layer expands the dimension of the inputs.
• Figure 2: Photon detection systems in ProtoDUNE-like (left) and DUNE-like geometries (right). The grey rectangles are the photon detectors.
• Figure 3: Feature-learning capability. In each figure, the $x$ axis is the position of the light sources along specific directions: $x$, $y$, and $z$ direction, while the $y$ axis shows the photon detection probabilities (visibility), on particular photon detectors (PDs). The top two rows show the visibilities in ProtoDUNE-like geometry while the bottom two are for the DUNE-like geometry. The visibilities from Genn prediction are plotted on the first and third rows and those from Geant4 simulation are on the second and forth rows. Note: positions of particular PDs and light sources are showed in Table. \ref{['tab:pdpos']}.
• Figure 4: Discrepancy of total detected photons per scintillation vertex. Left: discrepancies variation along the $x$-axis. Right: overall discrepancies, where the red dashed line is a Gaussian fit. Comparison for ProtoDUNE-like geometry is on the top and DUNE-like geometry on the bottom.
• Figure 5: Distribution of the detected photons on photon detection system. Top: the sum of detected photons on photon detectors from 200 MeV monoenergetic muon events starting from one specific position in the ProtoDUNE-like geometry. The response is much lower around photon detector #40 is caused by the fact that those photon detectors are the segmented ones which are much smaller than the bar-like light collectors. Bottom: the sum of detected photons on photon detectors from supernova neutrinos uniformly distributed in the DUNE-like geometry.