MACHINE LEARNING LOCAL WALL STEAM CONDENSATION MODEL IN PRESENCE OF NON-CONDENSABLE FROM TUBE DATA

Abstract

During a severe accident in nuclear reactors, steam condensation on containment structures is an important phenomenon that may affect the local concentration of hydrogen and the location of flammable regions in the nuclear containment. Accurate predictions of steam condensation rates and thereby peak hydrogen concentrations, temperature, and pressure rise in containment require the use of computational fluid dynamics (CFD) tools. The popular regulatory CFD calculations require a local heat transfer coefficient (HTC) at small, discretized length scales. In a classical three-dimensional full CFD, the HTC requirement can be eliminated, but for large structures and finely resolved multiscale calculation it may not be possible. This paper presents the development of two different kinds of local condensation HTC models for tube-based geometry based on (i) the machine learning (ML) model and (ii) the conventional third-order polynomial regression model. An extensive literature review was utilized to collect the data from various open literature sources. This eliminates the limitations of individual correlations and gives a best optimized model, which is valid for a wide range of flow regimes and conditions as compared to a specific correlation. Application of bulk HTCs for smaller tube as a local wall HTC is explored. Various simple ML models are compared for their performance against test data, and a multivariate adaptive regression splines (MARS)-based model was finally adopted for application and detailed discussion. The present ML was developed on the Python language framework. The MARS model was compared against the data, which was not used for the training and conventional polynomial based correlation. For traditional containment safety applications, both models were found to be suitable based on present studies.