Determination of the CMSSM Parameters using Neural Networks

TL;DR

This paper demonstrates that neural networks can effectively determine CMSSM parameters from LHC measurements, outperforming traditional  minimization, with high accuracy for key parameters using simulated data.

Contribution

The study introduces a neural network approach to invert the parameter-to-observable relation in CMSSM, achieving high-precision parameter estimation from LHC data.

Findings

Neural networks reliably determine m_0 and m_{1/2} with errors as low as 1%.

The method outperforms  minimization in accuracy for parameter extraction.

High-precision estimates of tan and A_0 are possible with larger datasets.

Abstract

In most (weakly interacting) extensions of the Standard Model the relation mapping the parameter values onto experimentally measurable quantities can be computed (with some uncertainties), but the inverse relation is usually not known. In this paper we demonstrate the ability of artificial neural networks to find this unknown relation, by determining the unknown parameters of the constrained minimal supersymmetric extension of the Standard Model (CMSSM) from quantities that can be measured at the LHC. We expect that the method works also for many other new physics models. We compare its performance with the results of a straightforward \chi^2 minimization. We simulate LHC signals at a center of mass energy of 14 TeV at the hadron level. In this proof-of-concept study we do not explicitly simulate Standard Model backgrounds, but apply cuts that have been shown to enhance the signal-to-background ratio. We analyze four different benchmark points that lie just beyond current lower limits on superparticle masses, each of which leads to around 1000 events after cuts for an integrated luminosity of 10 fb^{-1}. We use up to 84 observables, most of which are counting observables; we do not attempt to directly reconstruct (differences of) masses from kinematic edges or kinks of distributions. We nevertheless find that m_0 and m_{1/2} can be determined reliably, with errors as small as 1% in some cases. With 500 fb^{-1} of data tan\beta as well as A_0 can also be determined quite accurately. For comparable computational effort the \chi^2 minimization yielded much worse results.