Multibody simulation (MBS) is an established method for modeling wind turbines for purposes such as load calculation. While capable of modeling wind turbines at high levels of detail, large MBS models can be computationally expensive. This makes deployment for real-time critical use cases like digital twins difficult. Machine learning models can be a faster alternative because they only capture relevant details and scale well with parallelization. In this paper, a real-time capable surrogate model is developed with the goal of capturing the operating mode and the tower oscillations of an MBS model. Firstly, several machine learning models based on different architectures are being scaled so that they can be evaluated within the available computational budget. The selected models are being trained on the MBS results. The trained models are being run in a loop with the turbine’s controller and an optional numerical tower oscillator to create the surrogate model. All tested models show good convergence behavior and high R 2 values during training. However, running the models in the surrogate environment proves more difficult as some models diverge from the expected results. The best performing model in the surrogate environment is determined using quantitative criteria.