Abstract : The world’s very first exascale computer has finally arrived. The first generation of exascale machines will predominantly rely on hybrid architectures where massive numbers of CPUs, GPUs, and specialized hardware accelerators, coexist. Realizing the full potential of such architectures is a formidable task that requires an in-depth rethinking of current approaches. In this mini-symposium, we address the challenges faced by computational materials and chemical science communities. We specifically explore novel techniques, algorithms, and methodologies that can extend the time and length scales of atomistic simulations using exascale hardware.
Organizer(s) : Joshua Finkelstein, Danny Perez, Emanuel Rubensson, Tony Lelièvre
[01608] Quantum Materials Simulations at the Nexus of Exascale Computing, Artificial Intelligence, and Quantum Computing
Format : Talk at Waseda University
Author(s) :
Aiichiro Nakano (University of Southern CaliforniaUniversity of Southern California)
Abstract : Computing landscape is evolving rapidly. Exascale computers have arrived, and quantum supremacy has been demonstrated for several problems, while artificial intelligence (AI) is transforming every aspect of science and engineering. Atomistic simulations at the exa-quantum-AI nexus are revolutionizing quantum materials research. I will describe research and education on atomically thin two-dimensional and other quantum materials using our AI and quantum-computing enabled exascale materials simulator (AIQ-XMaS). Specifically, I will describe (1) self-assembly of layered material metastructures for scalable and robust manufacturing of quantum emitters for future quantum information science and technology; and (2) excited-state neural-network quantum molecular dynamics (NNQMD) trained by first-principles nonadiabatic quantum molecular dynamics (NAQMD) to prove the exciting concept of picosecond optical, electrical and mechanical control of symmetric breaking in topological ferroelectric skyrmion and skyrmionium for emerging ultralow-power polar topotronics. This research was supported by NSF Future Manufacturing Program, Award 2036359, NSF Cybertraining Program, Award 2118061, and Sony Research Award. Simulations were performed at Argonne Leadership Computing Facility under DOE INCITE and Aurora Early Science programs and at Center for Advanced Research Computing of the University of Southern California.
[01950] Large scale quantum chemistry with Tensor Processing Units
Format : Online Talk on Zoom
Author(s) :
Ryan Pederson (University of California, Irvine)
John Kozlowski (University of California, Irvine)
Ruyi Song (Duke University)
Jackson Beall (SandboxAQ)
Martin Ganahl (SandboxAQ)
Markus Hauru (Alan Turing Institute)
Adam Lewis (SandboxAQ)
Shrestha Basu Mallick (Google LLC)
Volker Blum (Duke University)
Guifre Vidal (Google LLC)
Abstract : We demonstrate the use of Googles cloud-based Tensor Processing Units $\text{(TPUs)}$ to accelerate and scale up conventional $\text{(cubic-scaling)}$ density functional theory $\text{(DFT)}$ calculations. Utilizing $512$ TPU cores, we accomplish the largest such DFT computation to date, with $247,848$ orbitals, corresponding to a cluster of $10,327$ water molecules with $103,270$ electrons, all treated explicitly. Our work thus paves the way toward accessible and systematic use of conventional DFT, free of any system-specific constraints, at unprecedented scales.
[02042] Quantum molecular dynamics using Tensor cores
Format : Talk at Waseda University
Author(s) :
Joshua Finkelstein (Los Alamos National Laboratory)
Emanuel H Rubensson (Uppsala)
Susan M Mniszewski (Los Alamos National Laboratory)
Christian F. A. Negre (Los Alamos National Laboratory)
Anders M Niklasson (Los Alamos National Laboratory)
Abstract : Tensor cores represent a new form of hardware acceleration specifically designed for deep neural network calculations. They provide extraordinary speed and efficiency but were designed for low-precision tensor contractions. Despite this, we demonstrate how Tensor cores can be applied with high efficiency to the challenging and numerically sensitive problem of quantum-based Born–Oppenheimer molecular dynamics, which requires highly accurate electronic structure optimizations and conservative force evaluations.
[02949] Recent algorithmic improvements in parallel long-time molecular dynamics
Format : Talk at Waseda University
Author(s) :
Danny Perez (Los Alamos National Laboratory)
Abstract : The temporal reach of molecular dynamics is limited by poor parallel strong-scaling: even on exascale computers, direct simulation is expected to remain limited to microseconds. We explore alternative time-wise parallelization schemes that target long-time simulations where multiple trajectory segments are evolved simultaneously and assembled into a unique dynamically correct trajectory. We discuss recent speculative execution and resource-allocation schemes that suggest significant potential scalability enhancements, leading to increased simulation timescales when deployed on massively-parallel computers.
[03771] Adaptive parareal method for the simulation of atomistic defects
Format : Online Talk on Zoom
Author(s) :
Olga Gorynina
Tony Lelievre (Ecole des Ponts)
Frederic Legoll (Ecole des Ponts)
Danny Perez (Los Alamos National Laboratory)
Abstract : Molecular dynamics simulations demand extensive computational resources to accurately calculate ensemble averages and dynamical quantities over long trajectories. In this study, we employ an adaptive parareal algorithm to enhance the speed of MD simulations by breaking down time calculations into smaller segments. We focus on the diffusion of a self-interstitial atom in a body-centered cubic tungsten lattice, utilizing LAMMPS molecular dynamics software and employing machine-learned spectral neighbor analysis potentials (SNAP) and embedded-atom method potentials (EAM). The adaptive parareal algorithm, which iteratively refines approximate solutions using parallel fine solvers, demonstrates significant computational gains. The goal of the talk is to highlight the potential of the adaptive parareal algorithm for accelerating MD simulations.
[04027] Fast, Accurate and Large-scale Ab-initio Calculations for Materials Modeling
Format : Talk at Waseda University
Author(s) :
Vikram Gavini (University of Michigan)
Sambit Das (University of Michigan)
Phani Motamarri (Indian Institute of Science)
Abstract : This talk will present our recent advances towards the development of computational methods and numerical algorithms for conducting fast and accurate large-scale DFT calculations using adaptive finite-element discretization, which form the basis for the recently released DFT-FE open-source code (https://github.com/dftfeDevelopers/dftfe). The computational efficiency, scalability and performance of DFT-FE will be presented. Some application problems that highlight the utility of DFT-FE in tackling complex aperiodic systems will be demonstrated.
[04848] The Chunks and Tasks Matrix Library
Format : Talk at Waseda University
Author(s) :
Emanuel Rubensson (Uppsala University)
Abstract : We present the Chunks and Tasks Matrix Library, which is a parallel sparse matrix library able to dynamically take advantage of data locality in matrices to avoid movement of data. The library uses a sparse quadtree representation of sparse matrices and is implemented using the Chunks and Tasks programming model. We demonstrate the scaling capabilities for operations used in large-scale electronic structure calculations, including sparse matrix-matrix multiplication and algorithms for inverse factorization.
[05282] From Langevin dynamics to kinetic Monte Carlo: mathematical foundations of accelerated dynamics algorithms
Format : Online Talk on Zoom
Author(s) :
Tony LELIEVRE (Ecole des Ponts ParisTech)
Abstract : We will discuss models used in classical molecular dynamics, and some mathematical questions raised by their simulations. In particular, we will present recent results on the connection between a metastable Markov process with values in a continuous state space (satisfying e.g. the Langevin or overdamped Langevin equation) and a jump Markov process with values in a discrete state space. This is useful to analyze and justify numerical methods which use the jump Markov process underlying a metastable dynamics as a support to efficiently sample the state-to-state dynamics (accelerated dynamics techniques à la A.F. Voter). It also provides a mathematical framework to justify the use of transition state theory and the Eyring-Kramers formula to build kinetic Monte Carlo or Markov state models.
References:
- G. Di Gesù, T. Lelièvre, D. Le Peutrec and B. Nectoux, Jump Markov models and transition state theory: the Quasi-Stationary Distribution approach, Faraday Discussion, 195, 2016.
- G. Di Gesù, T. Lelièvre, D. Le Peutrec et B. Nectoux, Sharp asymptotics of the first exit point density, Annals of PDE, 5(1), 2019.
- T. Lelièvre, Mathematical foundations of Accelerated Molecular Dynamics methods, In: W. Andreoni and S. Yip (Eds), Handbook of Materials Modeling, Springer, 2018.
- T. Lelièvre, D. Le Peutrec and B. Nectoux, Eyring-Kramers exit rates for the overdamped Langevin dynamics: the case with saddle points on the boundary, https://arxiv.org/abs/2207.09284.
[05409] Quantum-Mechanical Shadow Born-Oppenheimer Molecular Dynamics for Distributed Computing
Format : Talk at Waseda University
Author(s) :
Anders Niklasson (LANL)
Abstract : We present recent developments of quantum-mechanical shadow Born-Oppenheimer molecular dynamics for simulations with tens-of-thousands of atoms using distributed, linear scaling, graph-based electronic structure calculations [Negre, Wall, and Niklasson, J. Chem. Phys. 158, 074108 (2023)].
[05380] Compressing, resampling and forecasting atomic simulations with descriptor vectors
Format : Online Talk on Zoom
Author(s) :
Thomas D Swinburne (CNRS )
Abstract : We show atomic descriptor functions give a metric latent space for atomic simulations, where sparse snapshots can be interpolated and complex transitions (e.g. yeilding) can be linearly classified. When descriptors are unimodal, latent trajectories can be resampled and forecasted by a vector autoregressive model, with a Mahalanobis extrapolation grade. The approach is applied to challenging, large-scale simulations of dislocation plasticity. A strategy to optimise resources is proposed, maximising the estimated information yield of additional effort.
[05479] Structure modeling with large-scale DFT and machine-learning methods
Format : Talk at Waseda University
Author(s) :
Tsuyoshi Miyazaki (National Institute for Materials Science (NIMS))
Abstract : To overcome the size limitation of DFT calculations, we have developed a large-scale and linear-scaling DFT code CONQUEST. Using CONQUEST, we can investigate the atomic and electronic structures of large and complex materials containing many thousands of atoms. In this talk, I introduce the recent progress of CONQUEST, together with a newly proposed method to analyze the local atomic structures observed in the DFT-MD simulations of complex systems.
