Preheating with deep learning

TL;DR

This paper explores using deep learning, specifically CNN-LSTM models, to analyze and predict the late-time dynamics of preheating after inflation, reducing computational costs significantly.

Contribution

It introduces a novel application of CNN-LSTM time series analysis to model preheating dynamics, decreasing reliance on costly numerical simulations.

Findings

CNN-LSTM accurately predicts late-time particle flow patterns.

Deep learning reduces simulation time and computational resources.

Universal behavior of preheating dynamics confirmed by wave kinetic theory.

Abstract

We apply deep learning techniques to the late-time turbulent regime in a post-inflationary model where a real scalar inflaton field and the standard model Higgs doublet interact with renormalizable couplings between them. After inflation, the inflaton decays into the Higgs through a trilinear coupling and the Higgs field subsequently thermalizes with gauge bosons via its $SU(2)\times U(1)$ gauge interaction. Depending on the strength of the trilinear interaction and the Higgs self-coupling, the effective mass squared of Higgs can become negative, leading to the tachyonic production of Higgs particles. These produced Higgs particles would then share their energy with gauge bosons, potentially indicating thermalization. Since the model entails different non-perturbative effects, it is necessary to resort to numerical and semi-classical techniques. However, simulations require significant costs in terms of time and computational resources depending on the model used. Particularly, when $SU(2)$ gauge interactions are introduced, this becomes evident as the gauge field redistributes particle energies through rescattering processes, leading to an abundance of UV modes that disrupt simulation stability. This necessitates very small lattice spacings, resulting in exceedingly long simulation runtimes. Furthermore, the late-time behavior of preheating dynamics exhibits a universal form by wave kinetic theory. Therefore, we analyze patterns in the flow of particle numbers and predict future behavior using CNN-LSTM (Convolutional Neural Network combined with Long Short-Term Memory) time series analysis. In this way, we can reduce our dependence on simulations by orders of magnitude in terms of time and computational resources.