Computing Interface Curvature from Height Functions Using Machine Learning with a Symmetry-Preserving Approach for Two-Phase Simulations

Abstract

The volume of fluid (VOF) method is a popular technique for the direct numerical simulations of flows involving immiscible fluids. A discrete volume fraction field evolving in time represents the interface, in particular, to compute its geometric properties. The height function method (HF) is based on the volume fraction field, and its estimate of the interface curvature converges with second-order accuracy with grid refinement. Data-driven methods have been recently proposed as an alternative to computing the curvature, with particular consideration for a well-balanced input data set generation and symmetry preservation. In the present work, a two-layer feed-forward neural network is trained on an input data set generated from the height function data instead of the volume fraction field. The symmetries for rotations and reflections and the anti-symmetry for phase swapping have been considered to reduce the number of input parameters. The neural network can efficiently predict the local interface curvature by establishing a correlation between curvature and height function values. We compare the trained neural network to the standard height function method to assess its performance and robustness. However, it is worth noting that while the height function method scales perfectly with a quadratic slope, the machine learning prediction does not.