Improving MadGraph5 importance sampling efficiency by using neural networks

In this work, we present a two-staged unweighting method for Monte Carlo event generation. This method is based on the use of an Artificial neural network surrogate, which is trained to predict the events weights. The surrogate is used in a first unweighting step in order to avoid the evaluation of the true weights for the rejected events. Then, a second unweighting accounts for the error committed by the surrogate. This algorithm can accelerate the unweighting thanks to the much faster evaluation of the surrogate in respect to the evaluation of true weight, and to a small allowed overweight in the final sample. We test this algorithm for two scattering processes at 13 TeV, which are: pp → t ̄t and pp → e−e+ggd ̄d.