Ångstrom AI used Claude Code to run ~100,000 GPU jobs and beat Meta's UMA-OMC
Ångstrom AI (YC S24), working with the University of Cambridge and AstraZeneca, used Claude Code to autonomously drive an experiment loop on anycloud, running roughly 100,000 GPU jobs to train CSP-MACE-Å — a machine learning model that matches DFT accuracy at 10,000x the speed and outperforms Meta's UMA-OMC on crystal structure prediction benchmarks.
Score breakdown
CSP-MACE-Å is the first machine learning model to match DFT accuracy for crystal structure prediction while delivering a 10,000x speedup, and its training demonstrates that a Claude Code agent autonomously driving a cloud GPU experiment loop can replace much of the manual execution and bookkeeping in AI research workflows.
- 01CSP-MACE-Å is a machine learning interatomic potential developed by Ångstrom AI (YC S24) with the University of Cambridge and AstraZeneca.
- 02The model is 10,000x faster than DFT, reducing crystal structure calculations from weeks to minutes.
- 03CSP-MACE-Å is described as the first model to match DFT accuracy for crystal structure prediction.
Ångstrom AI (YC S24), together with the University of Cambridge (the Csanyi group) and AstraZeneca, released a paper presenting CSP-MACE-Å, a machine learning model designed to replace DFT in crystal structure prediction (CSP). DFT (density functional theory) is the quantum-mechanical gold standard for CSP but is extremely slow — calculations for a single molecule can take days to weeks. CSP-MACE-Å delivers the same accuracy as DFT at 10,000x the speed, reducing those calculations to minutes. It also outperformed Meta's UMA-OMC — the prior state-of-the-art machine learning interatomic potential for organic molecular crystals — across Ångstrom's and AstraZeneca's evaluation suites, and is described as the first model to demonstrate DFT-level accuracy for CSP.
To train CSP-MACE-Å, Ångstrom used Claude Code as an agent within their research iteration loop.
The stakes for pharmaceutical development are high: a single molecule can form multiple crystal structures (polymorphs) with different physical properties, and late-appearing polymorphs can render a drug unusable after distribution. The article cites the 1998 case of the HIV drug ritonavir, which had to be pulled and reformulated when a more stable crystal form appeared two years after market release, costing Abbott more than $250 million.
To train CSP-MACE-Å, Ångstrom used Claude Code as an agent within their research iteration loop. Researchers discussed experiment plans with Claude — which jobs to launch, which outputs to compare, which metrics to track — and Claude translated those plans into concrete actions: calling the anycloud CLI to launch GPU job batches, monitoring status, downloading results, and generating plots and summaries to inform the next hypothesis. The result was roughly 100,000 GPU jobs run almost entirely on multi-cloud spot instances across the team's own cloud accounts. The article notes that the same fan-out capability that accelerated the loop also introduced financial risk, as a misconfigured batch could generate thousands of dollars in spend before anyone noticed.
Key facts
- 01CSP-MACE-Å is a machine learning interatomic potential developed by Ångstrom AI (YC S24) with the University of Cambridge and AstraZeneca.
- 02The model is 10,000x faster than DFT, reducing crystal structure calculations from weeks to minutes.
- 03CSP-MACE-Å is described as the first model to match DFT accuracy for crystal structure prediction.
- 04It outperformed Meta's UMA-OMC — the prior state-of-the-art model for organic molecular crystals — on Ångstrom's and AstraZeneca's evaluation suites.
- 05Claude Code acted as an agent in the experiment loop, calling the anycloud CLI to launch jobs, monitor status, download results, and generate plots.
- 06The training run involved roughly 100,000 GPU jobs, almost entirely on multi-cloud spot instances.
- 07The article cites the 1998 ritonavir polymorph crisis, which cost Abbott more than $250 million, as motivation for rigorous crystal structure prediction.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 18, 2026 · 10:40 UTC. How this works →