Speaker
Description
In hadron colliders, event reconstruction of processes involving neutrinos is challenging because the initial partons cannot be fully constrained, detector responses are not ideal, and multiple interactions are overlaid in the same event.
Many algorithms have been proposed to reconstruct events with missing momentum information using additional constraints based on prior knowledge of specific physics processes, but these methods are difficult to generalize.
Recently, deep generative models have been explored to predict neutrino momentum by modeling its distribution, accounting for multiple solutions due to missing momentum information.
Existing generative models approximate the most probable predictions via multiple sampling steps, making inference computationally inefficient and potentially leading to inaccurate predictions.
We introduce Mixture Density Networks (MDNs) to model the distribution of possible neutrino momentum solutions as a mixture of Gaussians, where the distribution parameters are estimated using a novel latent attention network, enabling reliable predictions with a single inference step.
In this presentation, we demonstrate the performance of MDNs using simulations of the semi-leptonic $t\bar{t}$ process in proton-proton collisions at 14 TeV, evaluating both reconstruction accuracy and computational efficiency.
