High-Performance Simulation of Multicomponent Flows Using Regularized Lattice Boltzmann Method on GPUs

Resumo

We present a high-performance GPU-based framework for simulating multicomponent incompressible flows using a regularized and thread-safe Lattice Boltzmann Method (LBM). The model is compatible with diffuse interface dynamics between immiscible fluids and is built on a D3Q19 grid with a BGK collision operator. Regularization is carried out by means of a Hermite polynomial projection of the second-order non-equilibrium moments, effectively filtering out high-order non-hydrodynamic modes and increasing numerical stability in low-viscosity regimes. The simulation strategy adopts a thread-safe formulation which merges streaming and collision in a single step. This approach avoids race conditions and eliminates the need for double buffering schemes, resulting in a reduced memory footprint and better performance on GPU architectures. Two canonical, multicomponent three-dimensional test cases are explored: (1) an oscillating droplet and (2) a turbulent jet flow. Each fluid component is modeled with a dedicated particle distribution function — $f_i$ for mass and momentum, and $g_i$%  for the scalar phase field $\phi$. Interfacial dynamics are governed by phase gradients and curvature-driven surface tension forces derived from diffuse interface theory. The boundary conditions differ between the two configurations: the jet employs periodic contours along the side faces, an outflow condition at the top, and a prescribed inflow at the bottom of the domain; the droplet simulation uses fully periodic contours in all directions to emulate an unbounded environment. The simulations were carried out on a 256×256×1028 grid, on a single NVIDIA RTX 4090 consumer-level GPU. This benchmark highlights the efficiency of the method and confirms the ability of the thread-safe formulation to provide robust, high-performance calculations for multiphase LBM with crisp interface representation. The results are consistent with theoretical expectations and previous benchmarks. The proposed approach demonstrates strong potential for scalable and physically accurate simulations of multiphase and multicomponent flows, laying the groundwork for extensions to more complex geometries.