Computational Fluid Dynamics (CFD) is very important for improving manufacturing processes and increasing the efficiency within the steel industry. This study investigates the impact of flow fidelity on shape optimization using active learning techniques such as Bayesian optimization. The focus is on the Continuous Casting (CC) process, which is analyzed under two fidelity modes: first, considering single-phase transient flow, and second, integrating transient multiphase flow using the CFD and Multiphase Particle-in-Cell (MPPIC) method. In the CFD-MPPIC coupling, water is modeled using the Eulerian approach, while air bubbles are represented with the Lagrangian method. The design optimization of the submerged entry nozzle (SEN) in CC is crucial due to its influence on meniscus fluctuation levels. Higher fluctuation can lead to slag entrapment, impacting the quality of steel slabs. Therefore, SEN optimization is conducted using geometric seeding with an objective function based on height fluctuation ( \(\Delta \zeta \) ), derived from CFD results. A semi-open source coupling is introduced for optimizing the SEN using Bayesian optimization. The objective function \((\Delta \zeta _{\text {obj}})\) is formulated based on \(\Delta \zeta \) , which represents a local mesh point value derived from pressure data that varies with each time step in the simulation. Results show a significant 60% reduction in \(\Delta \zeta _{\text {obj}}\) compared to the baseline in the low-fidelity model (water), and a 12% reduction in the high-fidelity model (water-air). Optimized geometries from transient results are fed into the MPPIC model to assess fidelity effects with error percentage, with varying degrees of accuracy observed. This study demonstrates the efficacy of active learning techniques in optimizing SEN design, considering different levels of flow fidelity for CC.