Session F56: Harnessing Exascale Computing for Condensed Matter Simulations

The State of Exascale Quantum Monte Carlo

Ab initio Quantum Monte Carlo (QMC) methods currently offer the highest accuracy achievable for general materials. However, systematic developments are needed to extend their reach in terms of scale (atom and electron count), complexity of materials and physics, range of properties that can be computed and compared with experiment, and to address fundamental approximations such as the fixed-node approximation and Fermion sign problem. The combination of recently improved methods and algorithms plus increases in computation power promises to extend their reach to materials from across the entire periodic table and to enable the few underlying approximations to be systematically tested. I will present the current computational performance and scientific capabilities of real space QMC methods as implemented in the open source QMCPACK code[1,2]. Portability to and performance of Intel, AMD, and NVIDIA GPUs is achieved by using a careful software design[3] and state of the art software development approaches. Examples will be given of application to quantum materials such as TbMn₆Sn₆, MnBi₂Te₄, and 2D bilayers, where the combinations of electron correlation, charge, spin, and lattice couplings challenges more approximate electronic structure approaches and use of QMC is strongly merited. Finally, I will give an outlook of where progress is critically needed in the field.

1. P. R. C. Kent et al J. Chem. Phys. 152 174105 (2020), doi: 10.1063/5.0004860

2. J. Kim et al. J. Phys.: Condens. Matter 30 195901 (2018) doi: 10.1088/1361-648X/aab9c3

3. Y. Luo, P. Doak, and P. Kent, IEEE/ACM International Workshop on Hierarchical Parallelism for Exascale Computing (HiPar) (2022) 22 doi: 10.1109/HiPar56574.2022.00008   

Title: The prospect for molecular dynamics at the exascale

Molecular dynamics (MD) is one of the most powerful tools in the arsenal of computational materials scientists. While the exponential growth in available computing power has translated into proportional gains in terms of accessible simulation sizes and accuracies, similar gains have not been realized regarding simulation times, due to the breakdown of conventional domain decomposition approaches. I will review the main challenges facing the community with respect to fully realizing the promising of exascale computers and describe the progress made by the EXAALT project, funded under the DOE’s Exascale Computing Project. I will discuss advances in optimizing modern ML-based potentials for exascale architectures, in long-time simulations using parallel-in-time algorithm, and in the integration of ML workflows at scale.

    

Large-scale DFT calculations for nanoarchitectonics

In nanoarchitectonics, utilizing the interaction between nanoscale systems is important to create new material functionality. First-principles density functional theory (DFT) calculation is a powerful tool to investigate the correlation between the atomic and electronic structures which cause materials properties. To describe complex nanoscale structures such as interfaces, defect complexes and disordered systems, large-scale structural models consisting of several thousands of atoms or more are required. However, conventional DFT calculations are limited to be typically less than a thousand atoms because the computational cost scales cubically to the number of atoms N.

To overcome the limitation, we have developed a large-scale DFT code CONQUEST [1,2]. CONQUEST can treat large systems by using local orbitals to express density matrices and a linear-scaling (O(N)) method based on the density matrix minimization. The computational cost scales cubically to the number of the support functions, both in the O(N) and the conventional diagonalization calculations. Therefore, to reduce the number of support functions without losing accuracy, we have introduced multi-site support functions (MSSF) [3]. MSSFs are the linear combinations of pseudo-atomic orbitals from a target atom and its neighbor atoms in a cutoff region. MSSFs correspond to local molecular orbitals so that the number of required support functions can be the minimal.

We have also proposed an efficient and comprehensive method to find out the atoms with characteristic electronic structures in large-scale systems by using statistical methods. We have investigated supported metallic nanoparticles in which the interaction between the nanoparticle and the support base plays an important role for catalytic reactivity [4]. By our large-scale DFT calculation and the statistical analysis, the size- and site-dependence of atomic and electronic structures have been investigated for isolated and supported nanoparticles.   

The reach of exascale architectures across the physical sciences

The talk presents perspectives of scientific high-performance computing (HPC) in the exascale era, with a focus on the challenges and opportunities posed by the underlying hardware and software technologies to computational physics. We shall sketch selected activities towards enabling "grand challenge" science applications of the Max Planck Society on exascale-class supercomputers, highlighting the role of multidisciplinary application development teams composed of domain scientists, computer scientists, mathematicians, HPC experts, and of collaborations with hardware and software vendors. Taking the view of a computational physicist, we will specifically address some major technical and strategic programming challenges and share our experiences. opinions, and some practical recommendations for mastering them.   

Large-scale quantum-accurate atomistic simulation of plasma-facing materials for fusion energy

Plasma-facing materials within fusion reactors are subject to extreme conditions including high particle fluxes of a variety of plasma species and high heat loads.  Developing materials that can handle these extreme conditions is difficult but atomistic modeling like molecular dynamics (MD) can play a key role in fundamental understanding of how these materials degrade in high radiation environments.  However, one of the limitations of MD simulations is the lack of accurate interatomic potentials (IAPs) especially when modeling materials in extreme environments like at the plasma-material interface.  New machine learned interatomic potentials like the Spectral Neighbor Analysis Potential (SNAP) have been shown to have a higher accuracy compared to classical potentials, allowing for quantum accuracy with MD scalability.  We have developed a series of SNAP potentials for studying plasma-material interactions in W, W-ZrC, and MoNbTaTi.  SNAPs optimization on exascale machines allows for unprecedented extremely large, highly accurate MD simulations.  In this presentation, we will discuss the development of SNAP interatomic potentials for modeling plasma-material interactions and subsequent simulations of large-scale MD simulations investigating the effect of plasma exposure to candidate plasma-facing materials using leadership class computing resources.  SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.