Advancing Quantum Many-Body GW Calculations on Exascale Supercomputing Platforms

Understanding and designing advanced materials requires highly accurate computer simulations, but growing system complexity is driving up computational cost. Quantum many-body GW methods provide one of the most precise ways to model electronic excited states and the interactions between electrons, delivering accuracy beyond standard approaches. However, their large memory requirements and intensive communication patterns make them difficult to scale efficiently on modern supercomputers. A team of researchers from USC and Berkeley Lab implemented innovative GW capabilities within the BerkeleyGW software package, enabling large-scale simulations on exascale supercomputers including ALCF’s Aurora system. The team’s work was recognized as a finalist for the 2025 ACM Gordon Bell Prize.

Challenge

Quantum materials exhibit complex electronic and many-body phenomena that are critical for next-generation computing, energy, and electronics applications. Accurately predicting properties such as electron motion, excitations, and electron-phonon interactions requires methods beyond standard approximations like density functional theory (DFT). Traditional GW approaches, while highly accurate, are limited by computational cost and memory requirements, making simulations of large, heterogeneous systems such as solid-state defects or moiré superlattices extremely challenging. The team needed to develop methods that could scale across thousands of nodes and GPUs while maintaining predictive accuracy for both electronic excitations and dynamic interactions.

Approach

The researchers implemented multiple methodological and algorithmic innovations in the BerkeleyGW software package, including performance portability across heterogeneous GPU architectures and kernel-level optimizations with hardware-aware programming. They introduced GW perturbation theory (GWPT), which couples electron motion and phonon interactions within a single framework, enabling simulations of electron-phonon coupling phenomena critical for material transport, optical absorption, and quantum-state lifetimes. These advancements allowed scaling to full-system runs on Aurora, achieving over 1 exaflops of sustained performance and exceptional strong and weak scaling across thousands of nodes. Aurora’s large memory capacity and scalable architecture were critical for carrying out memory-intensive simulations of systems with tens of thousands of atoms, allowing the team to capture quantum effects in larger and more complex systems that would not have been possible on previous-generation systems. The improvements also ensure performance portability across different platforms.

Results

The team’s approach has demonstrated exceptional versatility for systems up to 17,574 atoms, capturing quantum effects previously out of reach for large-scale simulations. Runs on Aurora and other DOE systems achieved unprecedented predictive accuracy for electronic and optical properties, enabling simulations of complex materials dynamics at the electron-phonon level. This effort represents a major breakthrough in exascale-enabled quantum materials modeling. The team’s work was named a finalist for the 2025 ACM Gordon Bell Prize, recognizing both the computational performance and scientific impact of the project.

Impact

These advancements establish a new standard for at-scale quantum materials simulations, enabling the study of electronic excitations, electron-phonon coupling, and dynamic many-body interactions in previously intractable systems. The improvements in BerkeleyGW provide a portable, efficient framework for large heterogeneous systems, empowering researchers to design and understand next-generation electronic, optical, and quantum devices with unprecedented accuracy. By making these capabilities available to the broader scientific community, the team’s work paves the way for new discoveries and predictive modeling across the rapidly evolving field of quantum materials.