Conditional Wasserstein GAN for Simulating Neutrino Event Summaries using Incident Energy of Electron Neutrinos

TL;DR

This paper introduces a Conditional Wasserstein GAN model conditioned on neutrino energy to efficiently simulate complex electron neutrino interactions, matching traditional methods in accuracy but with less computational cost.

Contribution

The study develops a novel CW-GAN architecture that generates full multidimensional neutrino event data without variable reduction, improving simulation efficiency.

Findings

Generated samples match GENIE distributions statistically.

Successfully captured complex non-linear correlations.

Reduced computational overhead compared to traditional methods.

Abstract

Event simulation for electron neutrino interactions plays a foundational role in precision measurements in particle physics experiments, yet the computational demand of traditional Monte Carlo methods remains a significant challenge, especially for complete, high-dimensional event reconstruction. In this study, we present a generative model based on the Conditional Wasserstein Generative Adversarial Network (CW-GAN) framework. This architecture is conditioned on the input neutrino energy. It utilizes a Wasserstein loss function, stabilized by a gradient penalty, to learn the complex mapping from a latent space to structured kinematic data. Our model is tailored to replicate the full multidimensional kinematics of electron neutrino interactions as described by the GENIE event generator. Our focus is specifically on the Inverse Beta Decay (IBD-CC), Neutral Current (NC), and nue-e-elastic scattering processes (NuEElastic), spanning an energy window of 10-31 MeV. Our approach abandons variable reduction schemes and instead generates the entire summary ntuple, enabling holistic event-by-event modeling. Training is performed separately for each of the three interaction types, with rigorous convergence monitoring over 100-300 epochs per channel. We perform a rigorous quantitative validation against held-out GENIE test datasets. The generated samples demonstrate fidelity, reproducing the 1D marginal distributions for all kinematic variables with statistical compatibility, and successfully capturing the complex non-linear correlations between them. This work offers a scalable and efficient alternative to traditional MC event generation, providing full-spectrum kinematic simulation for key electron neutrino interaction channels while drastically reducing computational overhead.