Preprint reports Kohn‑Sham DFT at 100–200 million atoms and an 11.3M-atom battery interface

A new arXiv preprint reports a version of Kohn-Sham density functional theory, or DFT, that the authors say scaled to simulations of 100 million and 200 million atoms, a size far beyond the range usually associated with first-principles materials modeling. In the same paper, the researchers say they used the method on an 11,325,600-atom model of a solid-state battery interface and obtained results that matched spectroscopy measurements. The work is a preprint, not a peer-reviewed journal article, and its performance and record claims are self-reported.

That caveat matters because the scale claimed here is unusual for DFT, a standard quantum-mechanical tool used to calculate the properties of materials from first principles. Conventional Kohn-Sham DFT is valued for its accuracy, but it typically becomes much more expensive as systems grow, often scaling roughly cubically in computation and quadratically in memory. In practice, that has long limited many routine calculations to hundreds or thousands of atoms, not the tens of millions needed to directly model experimentally sized interfaces.

The preprint, “Extreme-Scale Linear-Scaling Kohn-Sham DFT at 100 Million Atoms: Bridging Quantum Simulations and Experiments,” was posted to arXiv on Sept. 11, 2026. The 14-author team includes researchers affiliated with the National Supercomputing Center in Shenzhen, Peking University Shenzhen Graduate School, Sun Yat-sen University, Tsinghua University, the National Supercomputing Center in Wuxi and ETH Zurich.

The authors describe their framework, called XLSDFT, as a linear-scaling, or O(N), Kohn-Sham DFT method built around divide-and-conquer decomposition of the one-particle density matrix and Chebyshev-filtered subspace iteration. In plainer terms, the goal is to make the cost of the calculation rise roughly in proportion to system size rather than exploding as the model gets larger. “XLSDFT reduces computational complexity by orders of magnitude,” the authors wrote.

The largest benchmark runs were carried out on LineShine, the exascale supercomputer at the National Supercomputing Center in Shenzhen. For the 100 million-atom silicon scaling study, the paper reports 96.6% weak-scaling efficiency — meaning the code kept most of its efficiency as more computing resources were added to handle a proportionally larger problem — and 157.9 petaflops of sustained double-precision performance. Those figures are reported by the authors in the preprint and, based on the public record, have not been independently verified by a third party.

The paper’s most consequential example may be the battery simulation, because it is presented as more than a supercomputing milestone. The model was a Li/LGPS/Li interface, with LGPS referring to the solid electrolyte lithium germanium phosphorus sulfide, Li10GeP2S12. The preprint gives the simulation cell as roughly 90 by 44 by 57 nanometers and says it contained 11,325,600 atoms.

According to the authors, the calculation captured how lithium metal reacts with the solid electrolyte at atomic resolution. They report that computed observables including projected density of states and Bader charges were in quantitative agreement with X-ray photoelectron spectroscopy, or XPS, depth-profile measurements. The paper specifically says low-valence germanium and phosphorus were enriched at the interface and that the calculated depth trends matched the XPS data. The authors characterize that system size as roughly 1,000 times beyond prior DFT scale for such complex interfacial systems.

For historical context, the preprint says its 200 million-atom silicon calculation is about 20 times larger than the previous 10 million-atom DFT record, which it identifies as LS3DF’s 2024 silicon result.

The paper is marked as submitted to SC26, the major high-performance computing conference, and is listed in the ACM Gordon Bell finalists program for presentation on Nov. 17, 2026. But for now, the result should be read as a notable preprint claim: a reported attempt to push quantum-accurate materials simulation toward experimentally relevant length scales, pending peer review and further scrutiny.