- Tusas: A modern computational framework for microstructure evolution toward exascale
The Phase-field method is one of the most powerful techniques to simulate microstructure formation in metal alloys. Software development for phase-field simulation is an emerging need for multi-physics materials modeling capability. Also, the software will be highly impactful in this high-throughput, data-driven era with the ultimate goal of reproducible research and code reuse. We developed a fully-coupled, fully-implicit approach for phase-field modeling of solidification dynamics and microstructure formation in metals and alloys. Our simulation approach consisted of a finite element spatial discretization of the fully-coupled nonlinear system of partial differential equations treated implicitly by the Jacobian-free Newton-Krylov method with physics-based preconditioning. This approach allowed us to use timesteps larger than those restricted by the traditional explicit Courant-Friedrichs-Lewy maximum timestep limit. Also, our approach was algorithmically scalable due to an effective preconditioning strategy based on algebraic multigrid and block factorization. We implemented this approach by introducing a new open-source phase-field simulation framework, Tusas, developed at Los Alamos National Laboratory. We analyzed the numerical performance of Tusas in terms of algorithmic scalability while demonstrating the computational performance in terms of parallel scalability on emerging heterogeneous supercomputing architectures (right figure). We have demonstrated ideal strong and weak scaling on up to a billion unknowns and thousands of GPUs on the two fastest supercomputers (summit and sierra, right figure) in the USA. In addition, we demonstrated the efficacy of Tusas using benchmark phase-field simulations of dendritic solidification in metal alloys under additive manufacturing conditions (left figures, 2D and 3D dendrites). The benchmark simulations were validated extensively against related experimental measurements and physics-based analytical model predictions (middle figure). Currently, Tusas plays a critical role in studying solidification dynamics and microstructure formation during metal casting and additive manufacturing. Tusas is already available through a GitHub repository https://github.com/chrisknewman/tusas. Refer to publication article #21 for more details.