A Bayesian optimization-driven framework for CFD problems: Application to continuous casting

Advancements in Computational fluid dynamics (CFD) have enabled accurate simulation of complex physical phenomena encountered in industry. When combined with optimization techniques, CFD simulations can be used to optimize process parameters or structural geometries in production systems. Traditional intrusive optimization methods typically require extensive modifications to CFD codes, significantly limiting their applicability to complex problems. Among the non-intrusive alternatives, response surface methodology (RSM) approaches have gained increasing attention due to their versatility and broader applicability. However, RSM approaches often require a relatively large number of simulations, which is intractable for high-fidelity CFD models. Recently, Bayesian optimization (BO), an adaptive learning strategy that combines adaptive sampling with active learning principles, has emerged as a promising alternative that maximizes information gain from a limited number of strategically selected simulations. Based on this approach, the present work develops an efficient optimization framework for complex CFD problems, using exclusively open source tools. When benchmarked against an advanced optimization framework implemented in the commercial tool CAESES, the developed one demonstrates stability and consistency in identifying global optima. The framework is then applied to optimize the geometry of a submerged entry nozzle (SEN) for continuous casting, modeled using a single-phase water approximation. The results show that the proposed framework achieves optimization performance on par with its commercial counterpart, suggesting its applicability to a broad range of CFD-based engineering optimization.