Prediction of cavity length: Dimensionless group identification through neural network and active subspace method

Abstract

The prediction of cavity length is very important for identifying cavitation state. This paper introduces a sophisticated framework aimed at predicting cavity length, leveraging the combination of neural network architecture with the active subspace method. The model identifies the dominant dimensionless group influencing cavity length in hydrofoil and venturi. For hydrofoil, a linear, negatively correlated relationship is found between cavity length and its dominant dimensionless number. Conversely, for venturi, an exponential, positively correlated relationship is identified. Using the found dominant dimensionless number to predict the dimensionless cavity length, the average relative errors are 0.146 and 0.136, respectively. The expression of the dominant dimensionless number, combined with the input parameters, is simplified into structural and physical functions, thereby significantly reducing the dimensionality of input while increasing the average relative error to 0.338. This study enhances the understanding of data-driven cavitation features and offers guidance for cavitation control and prevention.