> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# DeepMind's Co-Scientist leaves simulation for the lab: 2D material synthesis, E. coli predictions, physician-vetted architecture
- URL: https://globalfeed.ai/en/deepminds-co-scientist-leaves-simulation-for-the-lab-2d-material-synthesis-e-coli-predictions-physician-vetted-architecture/
- Published: 2026-08-29T05:44:55.000Z
- Updated: 2026-08-30T15:27:04.000Z
- Description: The Gemini-based multi-agent system is validated in real experiments: it produced a lamellar 2D material via a semi-automated reactor, predicted E. coli swarming in line with unpublished measurements, and found an architecture beating six frontier models on HealthBench.
- Author: GlobalFeed Editor
- Tags: Google, DeepMind, Gemini, scientific research, dil-en, elle

Google DeepMind has published a paper moving Co-Scientist, its Gemini-based multi-agent system, out of simulation and into real experiments: the chain from hypothesis generation to experimentation and manuscript writing is validated with wet-lab results across several scientific domains.

In materials science, the system interfaced with a semi-automated chemical vapor deposition reactor to design an MXene precursor and produced a lamellar 2D material sharing key structural similarities with the Ti3C2Tx lattice; by adapting growth recipes to the lab's constraints it grew monolayer MoS2, MoSe2 and WS2 semiconductors in a single attempt. In biology, it predicted the swarming behaviour of engineered E. coli across IPTG gradients from sparse imaging data, and the predictions quantitatively matched unpublished wet-lab measurements.

On the computer-science side, the system discovered an inference-time scaling architecture that outperformed six frontier models on HealthBench Hard and Professional, and reduced potential clinical harm under blinded physician evaluation.

End-to-end generated papers went through a double-blind study with 30 domain experts across 450 reviews; reliability modules are reported to reduce hallucination and plagiarism. The 35-author paper was posted to arXiv on 27 August.

Source: [arXiv:2608.26701](https://arxiv.org/abs/2608.26701?ref=globalfeed.ai)