Learning to Inflate

TL;DR

This paper introduces a neural network-inspired method to generate inflationary potentials by optimizing their Taylor coefficients, leading to models that align well with observational data and exhibit complex multi-field dynamics.

Contribution

It presents a novel approach using gradient ascent on neural network weights to generate inflationary potentials with realistic phenomenology.

Findings

Models agree with experimental data for large-field local maxima.

Small-field inflection point models also match observations.

Two-field models develop significant inflaton curvature during descent.

Abstract

Motivated by machine learning, we introduce a novel method for randomly generating inflationary potentials. Namely, we treat the Taylor coefficients of the potential as weights in a single-layer neural network and use gradient ascent to maximize the number of e-folds of inflation. Inflationary potentials "learned" in this way develop a critical point, which is typically a local maximum but may also be an inflection point. We study the phenomenology of the models along the gradient ascent trajectory, finding substantial agreement with experiment for large-field local maximum models and small-field inflection point models. For two-field models of inflation, the potential eventually learns a genuine multi-field model in which the inflaton curves significantly during the course of its descent.