Unfolding an Atomistic World: Atomistic Simulation of Reactor Pressure Vessel Steel Across Year-and-Meter Scales
Atomistic simulation now scales to ten-quintillion atoms in 1.71 days per year.
A team of researchers from Chinese institutions, led by Kun Li, has unveiled AtomWorld, a groundbreaking atomistic simulation framework that models reactor pressure vessel (RPV) steel degradation across unprecedented spatial and temporal scales—from angstroms and picoseconds to meters and decades. This addresses the long-standing challenge of predicting RPV steel lifetime, which requires bridging atomistic mechanisms with engineering-scale service conditions. Previous methods relied on fitted degradation laws or limited atomistic kinetic Monte Carlo (AKMC) approaches that could not reach year-and-meter scales.
AtomWorld achieves this through three tightly coupled layers: a consequence-aware algorithm that learns state transitions over the ab initio energy landscape, a compute-dense and synchronization-light HPC pipeline co-designed with modern supercomputers, and a physically grounded voxel-parallel framework that extends simulation to engineering scales. The system models ten-quintillion atoms, completing one simulated service year in just 1.71 days. It runs on five leadership supercomputers with 92-97% scaling efficiency, peaking at 1.27 EFLOP/s (48% of the Lineshine system's FP64 peak). This paradigm shift enables direct atomistic modeling of real-world nuclear reactor degradation, potentially transforming safety assessments and material design.
- AtomWorld simulates ten-quintillion atoms across year-and-meter scales, a first for atomistic modeling.
- Achieves 1.27 EFLOP/s peak performance (48% of Lineshine peak) with 92-97% scaling efficiency across five supercomputers.
- Completes one simulated service year in 1.71 days using a consequence-aware AKMC algorithm and voxel-parallel framework.
Why It Matters
Enables direct atomistic prediction of nuclear reactor steel degradation, improving safety and material design at scale.