Extreme Scale Phase-Field Simulation of Sintering Processes

An improved grand-potential phase-field model of solid-state sintering for many particles

“…The solver for the model is implemented with finite differences within the Pace3D framework [13], based on the massively parallel and high-performance implementation of [14]. The problem of the degrees of freedom increasing with the number of PFs N is solved as in [14] by employing the locally reduced order parameter approach and only saving eight phases at most per computational point. A WENO-5 scheme [15] is employed for the calculation of the advection updates in order to reduce numerical diffusion.…”

Lustre I/O performance investigations on Hazel Hen: experiments and heuristics

“…The simulation framework is based on the phase-field method [15] and contains multi-physical coupling to fields such as temperature, concentration or stresses to include their effect on microstructure evolution. The solver contains modules for diffuse interface approaches (Allen-Cahn, Cahn-Hilliard) [15], grain growth [16][17][18], grain coarsening [19,20], sintering [21,22], solidification [23,24], mass and heat transport, fluid flow (Lattice-Boltzmann, Navier-Stokes) [25], mechanical deformation (elasticity, plasticity) [26][27][28], magnetism [29], electrochemistry [30] and wetting [31,32]. The equations are discretized in space with a finite difference scheme and in time with different time integration methods such as explicit Euler schemes and implicit Euler via conjugated gradient methods [22,33].…”

“…To efficiently compute the complex evolution equations, the solver is parallelized using domain decomposition based on MPI. Selected models are also manually vectorized and achieve up to 32.5 % single-core peak performance on the Hazel Hen as shown in [22]. Explicit kernels within the solver scale up to 98304 cores with 97% parallel efficiency [22].…”

“…Selected models are also manually vectorized and achieve up to 32.5 % single-core peak performance on the Hazel Hen as shown in [22]. Explicit kernels within the solver scale up to 98304 cores with 97% parallel efficiency [22]. In [22], early results of the present paper were used to massively improve I/O performance for large-scale simulations.…”

“…Explicit kernels within the solver scale up to 98304 cores with 97% parallel efficiency [22]. In [22], early results of the present paper were used to massively improve I/O performance for large-scale simulations.…”