Calomplification -- The Power of Generative Calorimeter Models

Abstract

Motivated by the high computational costs of classical simulations, machine-learned generative models can be extremely useful in particle physics and elsewhere. They become especially attractive when surrogate models can efficiently learn the underlying distribution, such that a generated sample outperforms a training sample of limited size. This kind of GANplification has been observed for simple Gaussian models. We show the same effect for a physics simulation, specifically photon showers in an electromagnetic calorimeter.

Event Generators for High-Energy Physics Experiments

“…In addition to driving progress in treating phase-space generation through deep networks, NN-generators have the structural advantage that they can be trained on any combination of simulated and/or measured events. They also provide an efficient way to define and ship standardized event samples, including a modest numerical advantage over a statistics-limited training dataset [467,468], and can be used for post-processing of standard simulations. In these functions, it is again crucial to understand and control the precision of these generative networks at the level needed for precision measurements [469,470].…”

Machine Learning and LHC Event Generation

“…Given the interpolation properties of neural networks and the benefits of their implicit bias in the applications described in Sec. 2, we can quantify the amplification of statistics-limited training data through generative networks [64,65].…”

“…Neural networks work much like a fit and not like an interpolation in the sense that they do not reproduce the training data faithfully and instead learn a smooth approximation [64,65]. This is where we can gain some intuition for a NN-uncertainty treatment.…”