Multiscale light-matter dynamics in quantum materials: from electrons to topological superlattices

Taufeq Mohammed Razakh, Thomas Linker, Ye Luo, Nariman Piroozan, S. J. Pennycook, Nabin Kumar, Albert Musaelian, Anders Johansson, Boris Kozinsky, Rajiv K. Kalia, Priya Vashishta, Fuyuki Shimojo, Shinnosuke Hattori, Ken‐ichi Nomura, Aiichiro NakanoarXiv 2025

Abstract

Light-matter dynamics in topological quantum materials enables ultralow-power, ultrafast devices. A challenge is simulating multiple field and particle equations for light, electrons, and atoms over vast spatiotemporal scales on Exaflop/s computers with increased heterogeneity and low-precision focus. We present a paradigm shift that solves the multiscale/multiphysics/heterogeneity challenge harnessing hardware heterogeneity and low-precision arithmetic. Divide-conquer-recombine algorithms divide the problem into not only spatial but also physical subproblems of small dynamic ranges and minimal mutual information, which are mapped onto best-characteristics-matching hardware units, while metamodel-space algebra minimizes communication and precision requirements. Using 60,000 GPUs of Aurora, DC-MESH (divide-and-conquer Maxwell-Ehrenfest-surface hopping) and XS-NNQMD (excited-state neural-network quantum molecular dynamics) modules of MLMD (multiscale light-matter dynamics) software were 152- and 3,780-times faster than the state-of-the-art for 15.4 million-electron and 1.23 trillion-atom PbTiO3 material, achieving 1.87 EFLOP/s for the former. This enabled the first study of light-induced switching of topological superlattices for future ferroelectric 'topotronics'.

The Challenge of Multiscale Simulation

Designing the next generation of "topotronics"—ultralow-power, ultrafast devices based on quantum materials—requires understanding how light interacts with electrons and atoms at incredibly small and fast scales. This task presents a twofold challenge. First, simulations must bridge attoseconds ($10^{-18}$ s) of electronic excitation with nanoseconds ($10^{-9}$ s) of material response, and angstroms of atomic motion with microns of device-level features. Second, modern exascale supercomputers like Aurora are highly heterogeneous, featuring massive arrays of GPUs optimized for low-precision AI arithmetic, whereas traditional physics simulations typically demand high-precision (FP64) calculations.

The MLMD Framework: A Metascalable Paradigm

To address these challenges, researchers introduced the Metascalable Multiscale/Multiphysics Paradigm, implemented through the MLMD (Multiscale Light-Matter Dynamics) software framework. This approach shifts the paradigm from simple spatial decomposition to a combined spatial and physical decomposition.

Overview of the MLMD framework, illustrating the multiscale/multiphysics decomposition and the Metamodel-Space Algebra (MSA) used to bridge different levels of theory and hardware units.

The framework is built on two primary modules:

  • DC-MESH (Divide-Conquer Maxwell-Ehrenfest-Surface Hopping): This module handles quantum dynamics using a Divide-Conquer-Recombine (DCR) approach to split the problem into spatial domains and physical subproblems.
  • XS-NNQMD (Excited-state Neural-Network Quantum Molecular Dynamics): This module utilizes deep learning (the Allegro equivariant neural network model) to replace expensive electronic wave function calculations, providing first-principles accuracy at a fraction of the computational cost.

Technical Innovations: MSA and Shadow Dynamics

A key innovation is Metamodel-Space Algebra (MSA), a mathematical framework that allows different physical subproblems to reside on the hardware units best suited for them (e.g., GPUs for local fields, CPUs for complex logic), minimizing slow data transfers. To handle hardware heterogeneity, the framework employs Shadow Dynamics, using a GPU-resident "proxy" to solve local field dynamics in lower precision (FP32/BF16), while complex chemical interactions are managed by the CPU.

Detailed schematic of the (a) divide-and-conquer domains, (b) the DC-MESH module utilizing GPU/CPU heterogeneity for electron dynamics, and (c) the XS-NNQMD module using neural networks for atomic motion.

Furthermore, the Allegro-Legato approach introduces "sharpness-aware minimization" to the AI models. This ensures that small prediction errors do not accumulate over long-duration simulations, which is a common failure mode in large-scale AI-driven molecular dynamics.

Key Results and Scientific Impact

The MLMD framework demonstrated exceptional performance on the Aurora supercomputer. Using 60,000 GPUs, the DC-MESH module achieved 1.873 EFLOP/s, representing over 100% of the theoretical FP64 peak through the strategic use of mixed precision. Compared to previous state-of-the-art (SOTA) methods, the framework was 152x faster for quantum dynamics (15.4 million electrons) and 3,780x faster for neural-network molecular dynamics (1.23 trillion atoms) in PbTiO3 systems.

This computational power enabled the first-ever first-principles study of light-induced switching in a ferroelectric skyrmion superlattice—a complex "whirlpool" of polarization that could serve as the basis for future high-speed memory. The team is currently validating these predicted "topotronic" switching behaviors through experiments at the Linac Coherent Light Source (LCLS) at Stanford.

Future Outlook

While the framework achieves nearly perfect weak scaling, future work will focus on improving "strong scaling" (speeding up fixed-size problems) at extreme node counts. Additionally, the metascalable DCR/MSA paradigm is designed to eventually offload the most difficult quantum subproblems to emerging Quantum Processing Units (QPUs), further pushing the boundaries of quantum material simulation.

Summary updated

Abstract

Light-matter dynamics in topological quantum materials enables ultralow-power, ultrafast devices. A challenge is simulating multiple field and particle equations for light, electrons, and atoms over vast spatiotemporal scales on Exaflop/s computers with increased heterogeneity and low-precision focus. We present a paradigm shift that solves the multiscale/multiphysics/heterogeneity challenge harnessing hardware heterogeneity and low-precision arithmetic. Divide-conquer-recombine algorithms divide the problem into not only spatial but also physical subproblems of small dynamic ranges and minimal mutual information, which are mapped onto best-characteristics-matching hardware units, while metamodel-space algebra minimizes communication and precision requirements. Using 60,000 GPUs of Aurora, DC-MESH (divide-and-conquer Maxwell-Ehrenfest-surface hopping) and XS-NNQMD (excited-state neural-network quantum molecular dynamics) modules of MLMD (multiscale light-matter dynamics) software were 152- and 3,780-times faster than the state-of-the-art for 15.4 million-electron and 1.23 trillion-atom PbTiO3 material, achieving 1.87 EFLOP/s for the former. This enabled the first study of light-induced switching of topological superlattices for future ferroelectric 'topotronics'.
BibTeX
@misc{razakh2025multiscale,
  author = {Taufeq Mohammed Razakh and Thomas Linker and Ye Luo and Nariman Piroozan and S. J. Pennycook and Nabin Kumar and Albert Musaelian and Anders Johansson and Boris Kozinsky and Rajiv K. Kalia and Priya Vashishta and Fuyuki Shimojo and Shinnosuke Hattori and Ken‐ichi Nomura and Aiichiro Nakano},
  title = {{Multiscale light-matter dynamics in quantum materials: from electrons to topological superlattices}},
  year = {2025},
  eprint = {2509.00966},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.00966},
  howpublished = {arXiv},
}
Figure