Physics-aware machine learning for computational fluid dynamics surrogate model to estimate ventilation performance

Despite substantial advances in numerical simulation techniques, constructing a real-time optimization framework with accurate and fast predictions remains challenging. The difficulty arises from significant computational costs required for estimating a response of complex simulation models. Physics-informed machine learning (PIML) models could be an efficient alternative to solving multiple partial differential equations when boundary conditions change. This study aims to introduce an optimization model combined with a PIML algorithm, called physics-aware recurrent convolutional network (PARC), to explore an optimal ventilation efficiency in a confined engine room space during shipbuilding. Sixty computational fluid dynamics simulations were conducted to generate mean age of air (MAA) field data, which were split into training (45 cases), validation (5 cases), and testing (10 cases) datasets. Engine room structures and fan configurations were incorporated into the PARC model through a shape descriptor neural network, while MAA data were used to train the PARC model. The PARC model accurately predicted the temporal evolution of the MAA field, capturing complex ventilation fan information with an average prediction error of 1.5% at the final time step. Furthermore, the trained PARC model was coupled with the Bayesian optimization (BO) to explore the optimal ventilation efficiency. The results indicated that the optimized fan configurations reduced MAA values by up to 4.5%. The PARC-BO integrated framework offers a rapid and effective method for identifying fan configurations to enhance ventilation efficiency. It has potential applications in various industrial settings requiring improved air quality, such as power plants and coal mines.