Background and Objective: Voids in continuously cast steel billets are formed due to shrinkage during the solidification phase. Hot forming processes, such as multi-pass rolling, are crucial industrial methods for effectively closing such voids. If not closed properly, these voids can reduce the useful life of the component and diminish its mechanical properties. This project aims to accurately predict the kinetics of void closure during these processes using machine learning (ML) models. Methodology: First, the ranges of the parameters influencing the kinetics of void closure are identified through macro-scale rolling simulations within a data-driven multiscale strategy. This is followed by the use of Representative Volume Elements (RVEs), where boundary conditions are replicated. A comprehensive dataset documenting the evolution of the void under various rolling conditions is generated using Forge® software, based on the Finite Element Methods (FEM). This dataset serves as the foundation for developing the ML models. Results: Recurrent Neural Network (RNN) models are specifically designed to predict critical void parameters after each rolling pass, as the state at each step is required to predict the subsequent one. These parameters include void volume and inertia terms that define the void’s geometry and orientation. This methodology provides a powerful and efficient optimization tool for industrial billet production, significantly reducing dependence on computationally intensive full-scale simulations. The proposed methodology has been applied to a wide variety of thermomechanical paths and the results in the context of multipass processes will be discussed.