Artificial neural network model for single-phase real gas ejectors

Abstract Ejector is a passive gas dynamic device used ubiquitously in engineering applications. Applications of ejectors have transited from traditional steam extraction to heat-driven gas compression in supercritical and trans-critical CO 2 power and refrigeration cycles. The non-ideal gas behavior in such applications adds to the existing lack of fundamental understanding of the complex gas dynamics within the ejector. A number of simplified 1-D physics-based low-fidelity models have been proposed in the literature. These models offer reasonably good predictions for most ideal gases such as air and superheated steam. However, significant deviations as high as 20% between the experimental measurements and the model are observed for real gases. On the other hand, CFD simulations for an ejector are not only computationally expensive but also require careful selection and tweaking of the turbulence model. The paper presents a data-driven artificial neural network (ANN) model to predict the critical parameters of a supersonic ejector. The model is trained using a comprehensive experimental database comprising of various working fluids and operating conditions. The framework uses a topology of seven dimensionless input parameters capturing geometry, operating conditions, and working fluid characteristics to predict two output parameters; a) entrainment ratio and b) operational regime of the ejector (critical or mixed mode of operation). The model relies on a network of 84 neurons in the hidden layer and uses a split ratio of 0.15 for training the dataset. The trained ANN model is able to predict the entrainment ratio within 10% accuracy, with the ability to classify the operational mode with an accuracy of 91%.