Abstract Numerical simulation of interface evolution in multiphase and multimaterial systems requires accurate estimation of geometric quantities such as interface normals and curvature, which govern the kinetics of the underlying physical processes. While several curvature estimation techniques exist, such as the height-function method, they generally require continuous volume fraction information and are not directly applicable to binarized fields, where volume fractions are represented by discrete values of 0 or 1. Such binary representations frequently arise in cellular automata (CA) simulations and segmented image data. In these cases, conventional counting-based approaches provide only a limited representation of the underlying interface geometry, often leading to reduced accuracy. In this work, we propose a machine-learning-based framework for curvature estimation in binarized fields as an alternative to traditional counting cell approaches. A neural network is first trained using continuous volume fraction fields generated from known interface geometries. The model is subsequently adapted to binarized fields through a transfer-learning strategy, and its performance is compared against a network trained directly on binarized data. The accuracy of the proposed approach is assessed using standard benchmark geometries, along with the estimates of bias originating from the simplification to binarized field data. Finally, the methodology is integrated into a CA interface-evolution framework to demonstrate curvature-driven microstructural evolution in an analytically verifiable single-grain setting and a multi-grain environment. The results demonstrate that the proposed approach provides accurate curvature estimates and reproduces the expected interface-evolution and grain-growth kinetics in binarized systems.
原文链接
分享
引用
