Google: ERA — AI system that automates scientific code writing
Google published ERA (Empirical Research Assistance) in Nature — a Gemini-powered system that uses tree search to evaluate thousands of computational approaches and automates the writing of expert scientific software. The Computational Discovery platform is already available to researchers.
This article was generated using artificial intelligence from primary sources.
On May 19, 2026, Google published a paper in Nature on the ERA (Empirical Research Assistance) system — an AI tool powered by Gemini models that automates the iterative process of writing and optimizing scientific code.
What is ERA and how does tree search work?
ERA addresses one of the fundamental bottlenecks in research: the endless repetition of computational experiments in search of the best approach. The system takes a scientific problem description and a success metric, then independently searches the literature, writes code, combines techniques, and evaluates results.
The key innovation is tree search optimization — a technique for systematic search that simultaneously evaluates thousands of possible solutions. Instead of linear testing, ERA builds a tree of possible approaches, identifies the most promising branches, and directs computational resources toward them. The result is expert-level code without the usual iterations a researcher would go through.
In which domains does ERA outperform existing models?
Demonstrations span six domains. Flu, COVID, and RSV infection forecasts generated by ERA consistently rank at the top of CDC public leaderboards, above classical ensemble models. In water resource management, the system produced significantly more accurate early estimates of spring runoff than California’s official Bulletin 120 report. ERA also developed atmospheric CO₂ models with spatiotemporal resolution not previously achieved, and on the economic side, retail forecasts matched commercial consensus estimates.
Computational Discovery — public access to ERA and AlphaEvolve
Alongside the Nature publication, Google launched Computational Discovery through Google Labs — an experimental platform combining ERA and AlphaEvolve technology. The platform is part of the broader Gemini for Science initiative and is available to researchers through a trusted tester program.
All demonstrations are accompanied by code and experiments published on GitHub in the form of eight manuscripts covering ERA’s application in real-world scenarios. Google positions ERA as infrastructure for accelerating computational discovery — the process by which computational models generate verifiable scientific hypotheses and solutions.
Frequently Asked Questions
- What is ERA and how does its tree search work?
- ERA (Empirical Research Assistance) is a Gemini-powered AI system that takes a scientific problem description and a success metric, then independently searches the literature, writes code, combines techniques, and evaluates results. Its tree search optimization simultaneously evaluates thousands of possible solutions, identifies the most promising branches, and directs computational resources toward them.
- In which scientific domains has ERA been demonstrated?
- ERA has been demonstrated in six domains — flu, COVID, and RSV infection forecasting (consistently ranking at the top of CDC public leaderboards), water resource management (more accurate spring runoff estimates than California's official Bulletin 120), atmospheric CO₂ models with unprecedented spatiotemporal resolution, and retail forecasting matching commercial consensus estimates.
- What is Computational Discovery and who can access it?
- Computational Discovery is an experimental platform launched through Google Labs that combines ERA and AlphaEvolve technology. It is part of the broader Gemini for Science initiative and available to researchers through a trusted tester program.