The intensive usage of Monte Carlo event generators in elementary particle physics makes it necessary to use computing capacities efficiently. Surrogate-based unweighting algorithms are suitable for this purpose. In this work, neural networks are trained with SHERPA-generated events of the process group Z + 5jets to predict event weights. Factorisation properties of QCD matrix elements are exploited. A new loss function adapted for use in surrogate-based unweighting is presented. With an adjusted training method and after a final hyperparameter optimisation, significant savings in required computing time were achieved. A mean effective gain factor of feff = 2.6 ± 1.0 could be obtained.