This study presents a machine learning-based surrogate modeling framework for optimizing ventilation efficiency in complex industrial environments, with a focus on ship engine rooms and coffer dams. Traditional computational fluid dynamics (CFD) approaches are computationally intensive, particularly for large-scale parametric studies. To address this challenge, we employ a physics-aware recurrent convolutional (PARC) neural network, trained on 60 unsteady 3D CFD simulations with varying fan placements and orientations. The model predicts the Age of Air (AoA) distribution, a key indicator of ventilation performance, with high accuracy and drastically reduced computation time. The framework integrates geometric and operational parameters, such as fan location, orientation, and type, using a specialized shape descriptor. To identify optimal fan configurations that minimize AoA, Bayesian optimization is employed. The surrogate model achieves prediction times reduced from several hours to seconds, demonstrating its capability to replace high-fidelity CFD simulations in real-time optimization scenarios. Furthermore, the PARC model was evaluated on a different geometry, i.e., a coffer dam scenario, and showed successful generalization across different domains. In this case, we also implemented and compared two advanced operator learning models: the deep operator network (DeepONet) and the Fourier neural operator (FNO). These comparisons highlight the advantages and limitations of different operator-based approaches. Overall, this research demonstrates the potential of operator learning-based AI to accelerate fluid simulation workflows and support real-time digital twin applications in engineering design.

暂无