A compositional dynamic simulation model is fully implicitly integrated with a gas injection surface network model, to study the effects of CO2 injection into a depleted gas field. Multiple prediction scenarios are evaluated, under uncertainty, to reduce risk and improve decision making. We propose a workflow, composed of a geological sensitivity clustering step followed by a dynamic calibration step. The aim of this workflow is to decrease the objective function and improve the reliability of a probabilistic forecast, to model the CO2 storage potential of an onshore depleted gas field. Each run, containing all parameters and its objective function was exported and introduced into an inhouse R Script. Within this script we train a random forest tree to predict the objective function for various parameter combinations. This random forest is then used to generate 1 million models with the initial distribution from the simulation runs and will predict their objective function. The idea here is to get to a posterior distribution that can be used in the second simulation iteration. This method achieves a better history match within the ensemble, in a vastly reduced timeframe. History matched models were taken forward to predict CO2 injectivity. Injection variables, facilities and well completions for several wells have been included in the analysis, and numerical reservoir simulation models have been integrated with a surface network.