On May 19, 2026, Google Research’s ERA (Empirical Research Assistance) formally landed in Nature, with a title that states its ambition plainly: “AI system designed to help scientists write expert-level empirical software.” ERA uses Gemini to write and optimize scientific code, and the paper reports expert-level performance across six benchmark domains — genomics, public health, satellite imagery analysis, neuroscience prediction, time-series forecasting, and mathematics.
The accompanying Google Research post also announced eight manuscripts applying ERA to real scientific problems (five of them new), an open source release, and the gradual opening of Computational Discovery — built with both ERA and AlphaEvolve — through Gemini for Science. For research teams, this is a step toward turning “the research assistant who writes empirical code better than anyone” into infrastructure.
What ERA Is: Gemini Writes Scientific Code
One of the most time-consuming parts of empirical science is the loop: vary the data, adjust the model, rerun, compare, repeat. ERA automates that loop. Given a scientific problem and a success measure, it searches the literature, writes code, explores alternative solutions, combines existing techniques, and evaluates results — optimizing its output against the goal through a tree search that considers “thousands of options.” The algorithm was developed under Eser Aygun, Gheorghe Comanici, and Shibl Mourad. The design and performance first appeared as a preprint in fall 2025, followed by roughly six months of intensive experimentation before the Nature paper.
Expert-Level Results Across Six Domains
The Nature evaluation spans genomics, public health, satellite imagery analysis, neuroscience prediction, time-series forecasting, and mathematics, and ERA reached expert-level performance in all six. The subtext: expert-grade computational modeling no longer depends on whether your team happens to include someone who can write this kind of code.
Five New Applications: From Flu Forecasts to Solar Design
The five new manuscripts show how wide the applicability runs:
- Epidemiological forecasting of US hospital admissions at the state level, up to four weeks ahead, for flu, COVID-19, and RSV — consistently at or near the top of public CDC leaderboards
- Seasonal runoff forecasting for California’s snow-fed river basins, with early predictions significantly more accurate than the state’s official Bulletin 120 outlook
- Atmospheric CO2 mapping that combines GOES-East weather satellite data with other inputs to estimate emissions every 10 minutes, everywhere — capturing urban plumes over the Los Angeles basin and daytime plant uptake
- 3D solar energy collection optimized with Google Antigravity, where ERA found a 500-triangle volumetric fan that traps scattered radiation with zero backward shading
- Retail forecasting built on US economic indicators, Google Trends, and historical patterns, meeting or exceeding Action Economics’ consensus estimates and the Chicago Fed’s CARTS forecasts
External collaborators include epidemiologist Nicholas Reich and climate scientists Tapio Schneider, Ignacio Lopez-Gomez, and Anna Michalak, among others — ERA is positioned to amplify domain experts, not replace them.
Why This Is More Than Code Generation
The difference between ERA and a general code generator is the objective function. ERA is not optimizing for “looks like runnable code”; it optimizes for empirical performance on a stated scientific problem, measured against the success metric the researcher supplies. That makes it closer to a virtual postdoc: literature search, method combination, experiment design, and evaluation all sit inside one loop, and the loop runs until the metric improves. General-purpose coding assistants can draft a script from a prompt; ERA searches a solution space of thousands of candidate programs and keeps the ones that actually predict hospital admissions or river runoff better.
Google’s own framing is that this could democratize access to expert-level computational modeling — a real capability lift for regional research institutes and public-sector teams that lack software engineering capacity.
Open Source and What Opens Next
ERA’s code and experiments are open source on GitHub (google-research/era). Next comes Computational Discovery, which pairs ERA with AlphaEvolve and is gradually opening through a trusted tester program under Gemini for Science in Google Labs. It complements two sibling experiments — Hypothesis Generation, built with AI Co-Scientist and also published in Nature the same day, and Literature Insights — each covering a different stage of the scientific method. Research teams can register interest at the science page on labs.google now.
Sources
- Empirical Research Assistance (ERA): From Nature publication to catalyzing Computational Discovery — Google Research
- AI system designed to help scientists write expert-level empirical software — Nature
- Gemini for Science at I/O 2026 — Google Blog
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.