Solar wind forecasting is a core component of Space Weather, a field that has been the target of many novel machine-learning approaches. The continuous monitoring of the Sun has provided an ever-growing ensemble of observations, facilitating the development of forecasting models that predict solar wind properties on Earth and other celestial objects within the solar system. This enables us to prepare for and mitigate the effects of solar wind-related events on Earth and space.

The performance of some simulation-based solar wind models depends heavily on the quality of the initial guesses used as initial conditions. This work focuses on improving the accuracy of these initial conditions by employing a Recurrent Neural Network model. The study’s findings confirmed that Recurrent Neural Networks can generate better initial guesses for the simulations, resulting in faster and more stable simulations. In our experiments, when we used predicted initial conditions, simulations ran an average of 1.08 times faster, with a statistically significant improvement and reduced amplitude transients. These results suggest that the improved initial conditions enhance the numerical robustness of the model and enable a more moderate integration time step.

Despite the modest improvement in simulation convergence time, the Recurrent Neural Networks model’s reusability without retraining remains valuable. With simulations lasting up to 12 h, an 8% gain equals one hour saved per simulation. Moreover, the generated profiles closely match the simulator’s, making them suitable for applications with less demanding physical accuracy.