MadMiner: Machine Learning-Based Inference for Particle Physics

Abstract

The legacy measurements of the LHC will require analyzing high-dimensional event data for subtle kinematic signatures, which is challenging for established analysis methods. Recently, a powerful family of multivariate inference techniques that leverage both matrix element information and machine learning has been developed. This approach neither requires the reduction of high-dimensional data to summary statistics nor any simplifications to the underlying physics or detector response. In this paper we introduce MadMiner, a Python module that streamlines the steps involved in this procedure. Wrapping around MadGraph5_aMC and Pythia 8, it supports almost any physics process and model. To aid phenomenological studies, the tool also wraps around Delphes 3, though it is extendable to a full Geant4based detector simulation. We demonstrate the use of MadMiner in an example analysis of dimension-six operators in ttH production, finding that the new techniques substantially increase the sensitivity to new physics.

Benchmarking simplified template cross sections in W H production

“…More generally, the full exclusion limits based on the Sally method differ from those obtained from the linearized Fisher information, indicating the importance of squared new physics effects in this region of parameter space. In these parameter-space regions we expect that even stronger limits can be constructed with techniques that estimate the full likelihood function to all orders in the Wilson coefficients [23][24][25][26]. Closer to the SM, we find that the full limits are well approximated by the Fisher information, confirming that the linear operator effects dominate there.…”

“…We compared the STXS to a machine-learning-based analysis of the full, high-dimensional final state, using the Sally technique of Refs. [23][24][25] implemented in MadMiner [26] to calculate the statistically optimal observables and the maximal new physics reach of an analysis.…”

“…We use the implementation of the Sally algorithm in MadMiner 0.5 [26] to extract matrix-element information from our Monte-Carlo simulations, train neural networks to estimate the score, and calculate the expected limits on the Wilson coefficients. In addition to the four-momentum components, we choose to include a number of higher-level observables to aid in the training of the network, like the p T and η of each particle, the ∆R, ∆φ, invariant mass, transverse mass, and transverse momentum of each particle pair, altogether 48 observables for our W H process.…”

“…This latent variable is then integrated out in the machine learning step alongside with all other unobservable latent variables, such as the longitudinal neutrino momenta, initial state quark flavours or additional unobserved particles in reducible backgrounds (such as the jets in top pair production). Reference [26] contains examples to validate the performance of MadMiner's machine learning approach in the case of either a single observable or a single process. In the following, we will consider another simplified scenario in which we can validate MadMiner's reach estimate in the presence of both a high-dimensional observable space and irreducible backgrounds.…”

“…Until now, a major limiting factor in this approach was that detector effects and invisible particles like neutrinos could not accurately be taken into account. This problem was recently solved with a technique that combines matrix element information and machine learning [23][24][25], automated in the MadMiner tool [26].…”