Simulation of the coupled navier-stokes-cahn-hilliard-heat transfer system via a U-shaped spatial-channel attention neural operator

초록

Efficiently simulating coupled Navier-Stokes-Cahn-Hilliard-heat transfer system remains a significant challenge. This paper presents a U-shaped spatial-channel attention neural operator (U-SCANO) designed for phase-field simulation of binary incompressible heat-conductive fluids. U-SCANO employs a U-shaped architecture to efficiently extract multi-level features during down sampling while reconstructing the original resolution through an upsampling operator. Additionally, it incorporates a multi-scale spatial-channel attention fusion module (MSCAFM) within skip connections, enhancing feature fusion and improving representational capacity. Harnessing the synergy of its U-shaped architecture and MSCAFM, U-SCANO effectively captures and processes temporal features for simulating partial differential equation-driven systems. To validate the model's ability to capture complex flow dynamics, we conduct tests on the Kelvin-Helmholtz and Rayleigh-Taylor instability datasets. The results demonstrate that U-SCANO achieves stability and efficiency in long-term and time-dependent simulations, highlighting its potential for predicting the evolution of complex systems.

키워드