Neural network-based interface reconstruction algorithm for two-phase fluid flow

In two-phase fluid flow, the vapor-liquid interface tends to behave as a curved shape under the influence of surface tension. Curve reconstruction favors improving the resolution of bubbles or droplets in numerical studies. Based on artificial neural network (ANN) and convolutional neural network (CNN), two curve reconstruction algorithms, namely CIR-ANN and CIR-CNN, are proposed in this study. Both algorithms can achieve high-precision prediction of the center and radius when reconstructing interfaces using a portion of a standard circle, especially for the CIR-CNN algorithm. A strict mass conservation strategy is also proposed to ensure the reliability of the neural network predictions. In comparison with interface reconstruction algorithms such as piecewise linear interface construction (PLIC) algorithm, efficient least squares volume-of-fluid interface reconstruction algorithm (ELVIRA), quadratic spline based interface (QUASI) reconstruction algorithm after the first interface correction, Circle-based Interface Reconstruction (CIR), and CIR-ANN algorithm, the proposed CIR-CNN demonstrates good advantages in static interface reconstruction, with average accuracy ratios of 48.40, 86.69, 26.34, 4.07 and 2.26. However, when capturing the bubble under a complex rotation and shear recovery flow field, the advantage of the proposed algorithms decreases due to the increased complexity of the fluid volume fraction distributions. Regarding the computational time cost of reconstructing random circular interfaces, the proposed algorithms achieve average reduction ratios of 4.16, 7.86, and 34.14, respectively, compared to the CIR, ELVIRA, and QUASI algorithms.