Event generation and statistical sampling for physics with deep generative models and a density information buffer

Abstract

Simulating nature and in particular processes in particle physics require expensive computations and sometimes would take much longer than scientists can afford. Here, we explore ways to a solution for this problem by investigating recent advances in generative modeling and present a study for the generation of events from a physical process with deep generative models. The simulation of physical processes requires not only the production of physical events, but to also ensure that these events occur with the correct frequencies. We investigate the feasibility of learning the event generation and the frequency of occurrence with several generative machine learning models to produce events like Monte Carlo generators. We study three processes: a simple two-body decay, the processes e+e− → Z → l+l− and $$pp\to t\bar{t}$$ p p → t t ¯ including the decay of the top quarks and a simulation of the detector response. By buffering density information of encoded Monte Carlo events given the encoder of a Variational Autoencoder we are able to construct a prior for the sampling of new events from the decoder that yields distributions that are in very good agreement with real Monte Carlo events and are generated several orders of magnitude faster. Applications of this work include generic density estimation and sampling, targeted event generation via a principal component analysis of encoded ground truth data, anomaly detection and more efficient importance sampling, e.g., for the phase space integration of matrix elements in quantum field theories.

A factorisation-aware Matrix element emulator

“…Since then neural networks and other machine learning techniques have proved useful in many other areas of the field. On the theory prediction side they have been used to improve the efficiency of Monte Carlo sampling [1][2][3][4][5], to accelerate the simulation of radiation within a jet [6][7][8], to streamline the processes of generation and unweighting of simulated event samples [9][10][11][12][13][14][15][16][17][18]. Closer to the experimental measurements they have also been used to emulate detector simulation [19][20][21][22], they can be used to perform unfolding [23] or correcting for detector effects [24], and perform pileup subtraction [23].…”

Rare and Different: Anomaly Scores from a combination of likelihood and out-of-distribution models to detect new physics at the LHC

“…Variational autoencoders (VAEs) [28] and β-VAEs [29] accomplish this by maximizing the evidence lower bound, and as a result enforce a structured latent space. This structure can be transformed into an anomaly score, quantified for instance as the radius from the center of a Gaussian latent space or by using an information density buffer [30]. However, such methods suffer from the likelihood gap due to the intractable nature of the real likelihood, as well as from the assumption that the decoder is unable to reconstruct events outside the manifold spanned by the training set.…”

“…An increasingly popular method for finding such anomalous signals is to use anomaly detection techniques derived from Deep Learning (DL), see e.g. [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19][20].…”

Optimising simulations for diphoton production at hadron colliders using amplitude neural networks

“…There has also been a large focus on using ML for other components of MC event generator simulations. Specifically, Generative Adversarial Networks (GANs) [35] are being applied to event generation [36][37][38][39][40][41][42][43][44][45][46][47], event unweighting [48,49] and subtraction [50], with recent works incorporating Bayesian methods for uncertainty estimation into these generative methods [51]. NN-based approaches (some of which also use GAN technology) applied to parton showering [52][53][54][55] and event reweighting [56] have also been developed.…”