Multi-scale modeling and machine-learning-assisted optimization of convex-roll end-of-solidification reduction in ultra-thick continuous-casting steel slabs

End-of-solidification (EOS) reduction is essential for improving the internal soundness of ultra-thick continuous-casting steel slabs, yet its application is often constrained by the trade-off between roll capacity and cracking safety. Here we present an integrated framework that combines multi-scale modeling with machine-learning-assisted optimization to resolve this conflict. A macroscale thermo-mechanical finite-element (FE) model provides the local thermo-mechanical histories, which are mapped to a microscale crystal-plasticity FE model with cohesive-zone elements (CPFEM–CZM) to mechanistically assess grain-boundary decohesion and screen crack risk. To enable rapid process design, lightweight surrogates—polynomial regression and an XGBoost classifier—are trained on the multiscale simulation database to predict roll force and central densification, and to provide a conservative crack-risk screener that prioritizes minimizing missed cracks (Recall ≈ 0.92), while the overall separability remains moderate (AUC ≈ 0.635). The surrogates are then embedded into a constrained multi-objective evolutionary algorithm (NSGA-II) to obtain Pareto-optimal reduction schedules. The optimization indicates that convex-roll reduction applied at central solid fractions of 0.4 and 0.8 offers a favorable balance between central densification and cracking safety, whereas reduction around 0.6 leads to markedly higher crack risk. This work provides a transferable, data-driven route to inverse-design safe operating windows for EOS reduction in continuous casting.