Abstract:
Light-matter dynamics in topological quantum materials could enable ultrafast (petahertz) and ultralow-energy (attojoule) computing and sensing devices toward sustainable AI-embedded future. A challenge is simulating multiple field and particle equations for light, electrons, and atoms over vast spatiotemporal scales. Meanwhile, high-performance computing is at a historic crossroads, where traditional modeling and simulation applications may not survive the increasing heterogeneity and low-precision focuses of hardware.
We have developed a divide-conquer-recombine (DCR)/metamodel-space-algebra (MSA) paradigm to solve the multiscale/multiphysics/heterogeneity/low-precision challenge harnessing hardware heterogeneity and hybrid-precision arithmetic. We have thereby developed a MLMD (multiscale light-matter dynamics) software composed of first-principles DC-MESH (divide-and- conquer Maxwell-Ehrenfest-surface hopping) module for nonadiabatic quantum dynamics (NAQMD) and AI-accelerated XS-NNQMD (excited-state neural-network quantum molecular dynamics) module to expand the spatiotemporal scales of NAQMD. Using 60,000 GPUs of the Aurora supercomputer at Argonne National Laboratory, the DC-MESH and XS-NNQMD modules achieved nearly perfect scalability with 1.87 Exaflop/s performance for the former, thus allowing the simulation of light-induced switching of topological superlattices for future ferroelectric ‘topotronics’.
This work suggests new algorithm-hardware co-design pathways at the nexus of post-exascale computing, quantum computing, and AI.
Bio:
Aiichiro Nakano is a Professor of Computer Science with joint appointments in Physics & Astronomy, Quantitative & Computational Biology, and Collaboratory for Advanced Computing and Simulations at the University of Southern California (USC). He has authored 500+ refereed articles in the areas of scalable scientific algorithms, scientific visualization and machine learning, and computational materials science. He is a recipient of the National Science Foundation Career Award, Okawa Foundation Faculty Research Award, U.S. Department of Energy Aurora Early Science Program Award, and Best Paper Awards at IEEE/ACM Supercomputing, IEEE Virtual Reality, IEEE PDSEC, and ACM HPCAsia. He is a Fellow of the American Physical Society.
Summary:
Changing computing landscape for science
Current Exascale architecture is not ready to continue scaling
Many different modeling/compute architectures are being explored
Electronic
Quantum
AI-assisted
Focus: molecular dynamics
MD
Reactive MD
Nonadiabatic quantum MD: electron + photon interactions
Simulations used to train neural network potential functions
Quantum MD
Density Functional Theory
Replace n-electro problems with N 1-electron problems (ecponential to order N3)
Still limited to few hundred atoms
Adiabatic dynamics approximation takes advantage of the fact that electrons are so fast that they change states and move instantaneously relative to the nucleus
New work on O(N) DFT leverages data locality
Focus: non-adiabatic quantum molecular dynamics
Photon-electron interactions
Photons are much faster than electrons and cause state transitions in electrons
Must model electron dynamics explicitly
Key opportunity: Attosecond light-mater dynamics
Can revolutionize electronics, may enable pettahertz circuits
Need simulations to model this phenomenon
DC-MESH: divide & conquer Maxwell + Ehrenfest + surface-hopping)
LFD: local field dynamics
QXMD: Quantum excitation MD
LFD is shorted time-scales, maps well to GPU
QXMD: longer time-scales, maps well to CPU
Explicit coordination algorithms between LFD and QXMD on hybrid architectures
New HPC architectures good for running these simulations
But: computers are becoming more heterogeneous and complex
Many functional units
Low-precision computing units, 8 bits and even fixed-point 8 bits
Need to choose problems carefully to be solvable in lower precision
Approach:
Divide-conquer-recombine (DCR): across different regions of space/time AND different physical sub-problems of different computational characteristics
Meta-model-space algebra (MSA): localizes problems in different computation units
DCR/MSA: divides work into different units that use different compute units with different precision
AI-enhanced Multiscale Simulation
Multiscale light-matter dynamics: non-adiabatic QMD boosted by neural emulation
DCR: spatial division
Space is divided into overlapping domains
In each domain local electronics are solved
Reintegrated into the global electron field via a multi-grid hierarchical solver
Globally sparse/ locally dense eigen solver
Band-space-domain decomposition onto parallel machine
DCR:
LFD: local, structured dynamics fit naturally onto GPU
QXMD: longer-distance dynamics that require more complex computation: fit well onto CPU
Interactions can be captured in low precision
Implementation Innovations:
Open standards for compute infra
Data/loop reordering, blocking, hierarchical localization
GPU resident kernels
Parameterized mixed-precision computation
Reached Exaflop performance on light-matter dynamics
1.87 ExaFlop
Nearly perfect parallel efficiency
Neural net QMD:
Allegro-Legato
Allegro FM Foundation Model: many material properties & Vprocesses covering 89 elements
MSA3 - XN.NN - run neural model to model approximate dynamics, refine using full simulation
Opto-topotronics application
Next step: quantum dynamics on quantum computers
Exponentially hard topological quantum dynamics offloaded on to quantum computers
Next: what does “understanding” mean when AI does science?
Collaboration: cosmology, computer science, cinematic arts & philosophy (Gallow)
How do we understand physics