Photo-switching of a ferroelectric nanoscale structure in lead titanate. (Image: ALCF Visualization and Data Analytics Team.)
Understanding light-matter dynamics in topological quantum materials could help enable ultralow-power, ultrafast devices capable of meeting the growing demands of AI. A key challenge is simulating the coupled behavior of light, electrons, and atoms across vast spatial and temporal scales. These simulations must also run efficiently on exascale computers with increasingly heterogeneous architectures and lower-precision arithmetic. Using ALCF’s Aurora system, a team of researchers led by the University of Southern California achieved the first end-to-end exascale-deployed multiscale simulation of light-matter dynamics. Their approach incorporates machine learning to efficiently model the behavior of electrons and atoms, delivering significant improvements in time-to-solution. The team’s work was recognized as a finalist for the 2025 ACM Gordon Bell Prize.
Challenge
Light-matter interactions in quantum materials involve complex phenomena across atomic to device scales, combining electronic, atomic, and optical physics. Simulating these effects requires capturing ultrafast electron motion, atomic rearrangements, and electromagnetic fields in a single framework. Traditional approaches struggle to scale to systems with billions or trillions of atoms while maintaining high fidelity, making large-scale, high-precision simulations extremely challenging. The team needed to connect small-scale atomic motion to larger material structures and integrate AI to accelerate the calculations.
Approach
To tackle this problem, the researchers developed the Multiscale Light-Matter Dynamics (MLMD) software framework. It combines physics-based models of light, electrons, and atomic motion with AI to efficiently simulate extremely large systems. A key component is the Allegro-FM AI foundation model, which predicts atomic interactions across a wide range of materials, enabling simulations of over a trillion atoms with near quantum-level accuracy. The software also implements a divide-conquer-recombine approach, partitioning problems for optimal use of Aurora’s GPUs and CPUs, and then recombining the results to form a complete solution. Together, these innovations allow ultrafast, large-scale simulations that reveal how light can trigger structural changes in ferroelectric topological materials.
Results
Using 60,000 Aurora GPUs, the team’s MLMD software performed 152x and 3,780x faster than the state-of-the-art for 15.4 million-electron and 1.23 trillion-atom PbTiO₃ materials, respectively, achieving a sustained performance of 1.87 exaflops. This enabled the first study of light-induced switching of topological superlattices for future ferroelectric topotronics. The team’s achievement was recognized as a finalist for the 2025 ACM Gordon Bell Prize, highlighting both the scientific and computational breakthroughs enabled by AI-enhanced multiscale simulations.
Impact
This work demonstrates how AI and exascale computing can transform the study of complex quantum materials. By combining physics-based simulations with AI models like Allegro-FM, researchers can explore atomic-scale interactions in systems previously too large or complex to simulate. The framework provides a scalable approach for accelerating discovery in topotronics and other advanced electronic materials, generating insights to guide experiments and inform the design of ultrafast, low-power devices for next-generation computing technologies.
Publications
Kazakh, T. M., T. Linker, Y. Luo, N. Piroozan, J. Pennycook, N. Kumar, A. Musaelian, A. Johansson, B. Kozinsky, R. K. Kalia, P. Vashishta, F. Shimojo, S. Hattori, K. Nomura, and A. Nakano. “Multiscale Light-Matter Dynamics in Quantum Materials: From Electrons to Topological Superlattices,” SC25: The International Conference for High Performance Computing, Networking, Storage, and Analysis (November 2025), IEEE.
https://doi.org/10.1145/3712285.3771785
