A Differentiable Surrogate Model for the Generation of Radio Pulses from In-Ice Neutrino Interactions

TL;DR

This work proposes a modularized deep learning architecture to generate radio signals from in-ice neutrino interactions conditioned on the shower energy and viewing angle and ensures physical consistency of the samples and leads to advantageous computational properties when using the model as part of a bigger optimization pipeline.

Abstract

The planned IceCube-Gen2 radio neutrino detector at the South Pole will enhance the detection of cosmic ultra-high-energy neutrinos. It is crucial to utilize the available time until construction to optimize the detector design. A fully differentiable pipeline, from signal generation to detector response, would allow for the application of gradient descent techniques to explore the parameter space of the detector. In our work, we focus on the aspect of signal generation, and propose a modularized deep learning architecture to generate radio signals from in-ice neutrino interactions conditioned on the shower energy and viewing angle. The model is capable of generating differentiable signals with amplitudes spanning multiple orders of magnitude, as well as consistently producing signals corresponding to the same underlying event for different viewing angles. The modularized approach ensures physical consistency of the samples and leads to advantageous computational properties when using the model as part of a bigger optimization pipeline.

Figures (13)

• Figure 1: Figures reproduced from Barwick:2022vqt with permission. Left: A neutrino interaction in ice. The collision of a neutrino with an atom in the ice gives rise to a shower of secondary particles. These particles emit Askaryan radiation, which is strongest on the Cherenkov cone. In this plot, the viewing angle $\theta=\theta_C$, where $\theta_C$ denotes the Cherenkov angle in ice, and where the dotted red line indicates the direction towards the observer. Right: The same signal is depicted at different viewing angles, where $\Delta \Omega = \theta - \theta_C$. It can be observed that the signals with $\pm \Delta \Omega$ are approximately antisymmetric copies of each other.
• Figure 2: The modularized model architecture. The generator generates normalized samples at a fixed angle and for a given shower energy $E$. The $\theta$-Net then transforms the signal to the desired viewing angle $\theta$. Finally, the $a$-Net predicts the signal amplitude $a$ and combines it with the normalized signal $x_0$ to obtain the signal $x$.
• Figure 3: The U-Net architecture ronneberger2015u is often employed for transformations with the same input- and output size. It consists of multiple levels that allow for the extraction of features at different scales. Each level contains multiple ResNet blocks (the blue boxes). When moving up or down between levels, the outputs are down- or upsampled, respectively. Skip connections connect the different levels on both sides, from the encoder (left) to the decoder (right), enabling more stable learning. We employ U-Nets for the fine-tuning in the $\theta$-Net and the denoiser in the diffusion model.
• Figure 4: The architecture of the amplitude net ($a$-Net) is depicted. Features are first extracted from normalized signals and then combined with an embedding of $E$ and $\theta$. Then, these features are combined in a third network to yield log-amplitude predictions.
• Figure 5: The architecture of the $\theta$-Net. First, the normalized signals at the fixed reference angle $\theta_r=65.57^{\circ}$ are preprocessed via Equation \ref{['eq:preprocessing']}. A CNN is employed to extract additional features from the reference signal. Then, the preprocessed signals are finetuned using a U-Net, to yield normalized signals $x_0(\theta)$ at the desired angle $\theta$.