arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

BORN framework speeds up a large‑scale quantum‑chemical calculation: 1.41 h for a 50,024‑basis‑function Hartree‑Fock calculation | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Artificial IntelligenceEnglishPublished

BORN framework speeds up a large‑scale quantum‑chemical calculation: 1.41 h for a 50,024‑basis‑function Hartree‑Fock calculation

November 19, 2026arXiv: 2611.22897v1

BORN is a new computing framework that builds fast, high‑accuracy quantum‑chemical programs on the spot for each user’s problem.

The core idea is to look ahead. For a chosen calculation, BORN first creates a library of possible algorithmic designs from existing software. It then predicts each design’s error and how long it will run on the requested machine, using small test molecules and detailed micro‑benchmarks of the hardware. This foresight allows BORN to pick the best strategy that stays inside the user’s accuracy tolerance and to hand the right code to a coding agent that automatically compiles it.

Diagram showing BORN’s algorithmic stages: strategy space unfolding, error and runtime prediction, strategy selection, and agent‑driven compilation. (1/2)Diagram showing BORN’s algorithmic stages: strategy space unfolding, error and runtime prediction, strategy selection, and agent‑driven compilation. (2/2)
Figure 2. Pipeline illustration: BORN creates a strategy space, predicts target‑scale error and runtime, selects the fastest strategy within tolerance, and then compiles it into a task‑specific program.Han, Haozhi, Yang, Junjie, Zhang, Ruge, Huang, Zihe, Xu, Pengyi, Zhang, Bo et al. · arXiv:2611.22897v1 · Figure 2: · CC BY 4.0

In a demanding test on a molecule with 50,024 basis functions, this approach cut the runtime to 1.41 h on a single CPU node. The compact, compiled program was 77 times faster than a general‑purpose baseline that could not even run the calculation because its memory requirements exceeded 2^31. Including the nine‑hour development effort, the overall speedup was still nearly ten times. On a GPU, BORN finished the same calculation in 2.1 h—an order of magnitude faster than the only other available GPU‑accelerated package.

The results come from a careful study. The baseline model, tuned for the system, would have taken 109 h and required 700 GB of memory. BORN’s uncompiled code ran in 5.6 h and used 418 GB. After the agent‑driven compilation step, the final program ran in 1.41 h and used 334 GB, all while keeping the energy error below the user’s threshold.

Table summarizing runtime, memory usage, development time, and speedups for baseline tuned, Born uncompiled, and Born compiled scenarios.
Δ​ETotalMem.Dev.Speedup
Configuration(mEh)(h)(GB)(h)w/o Dev.w/ Dev.
Baseline, tuned†0.05210970001.0×1.0×
Born, uncompiled†0.0525.64183.819×12×
Born, compiled0.0531.413349.777×9.8×
Table 1. End‑to‑end runtime and memory comparison for a 50,024‑basis‑function Hartree–Fock calculation: baseline tuned (109 h, 700 GB), Born uncompiled (5.6 h, 418 GB), and Born compiled (1.41 h, 334 GB). The compiled Born program achieves a 77× speedup over the baseline and a 9.8× speedup when the nine‑hour development time is included.Han, Haozhi, Yang, Junjie, Zhang, Ruge, Huang, Zihe, Xu, Pengyi, Zhang, Bo et al. · arXiv:2611.22897v1 · Table 1: · CC BY 4.0

These findings show that, for large quantum‑chemical problems, BORN can deliver performance that would be impossible with conventional general‑purpose software alone. The method is limited by the accuracy of the error extrapolation and the hardware micro‑benchmarks, and it requires sufficient development resources to compile the chosen strategy. Nonetheless, the framework offers a clear path to rapid, reliable, system‑specific quantum‑chemical simulations.

Source paper

Born for the Task: Rethinking Quantum Chemistry with Foresighted Agentic Computing

Artificial IntelligencePhysics
View on arXiv