Summary: Nature published Google Research’s ERA on May 19, 2026, describing a Gemini-based system for producing and optimizing expert-level empirical scientific code. The evaluation covered genomics, public health, satellite imagery, neuroscience prediction, time-series forecasting, and mathematics, with expert-level results in each. Five new application manuscripts and an open-source code release accompanied the announcement.
Nature published Google Research’s ERA (Empirical Research Assistance) on May 19, 2026, under a title that leaves little to interpretation: “AI system designed to help scientists write expert-level empirical software.” ERA has Gemini do the writing, producing and optimizing scientific code, and the paper’s scoreboard covers six benchmark domains — genomics, public health, satellite imagery analysis, neuroscience prediction, time-series forecasting, and mathematics — with expert-level results in every one of them.
The companion Google Research post adds more: eight manuscripts that put ERA to work on real scientific problems (five of them released now for the first time), an open source code release, and the staged opening of Computational Discovery — built with both ERA and AlphaEvolve — through Gemini for Science. In effect, this is a step toward turning “the research assistant who writes empirical code better than anyone” into infrastructure.
How ERA Works: Giving Scientific Code to Gemini
Empirical science burns an enormous amount of time on one grinding cycle: swap the dataset, retune the model, rerun, compare, go again. ERA takes that cycle over completely. Hand it a scientific problem and a measure of success, and it will comb the literature, produce code, probe alternative solutions, stitch established techniques together, and score what comes back — steering itself toward the target with a tree search that weighs “thousands of options.” Eser Aygun, Gheorghe Comanici, and Shibl Mourad led development of the algorithm. The design and its performance were first made public as a preprint in fall 2025; roughly six months of intensive experimentation later, the work stood as a Nature paper.
Six Domains, Six Expert-Level Results
Each of the six domains in the Nature evaluation — genomics, public health, satellite imagery analysis, neuroscience prediction, time-series forecasting, and mathematics — ended with ERA judged expert-level. The message underneath the numbers: whether expert-grade computational modeling gets done no longer hinges on whether somebody on your team happens to be able to write this kind of code.
Five Fresh Applications, from Flu Forecasting to Solar Engineering
The five new manuscripts stretch across strikingly varied ground:
- State-level forecasting of US hospital admissions for flu, COVID-19, and RSV, reaching as far as four weeks ahead and holding steady at or near the top of the public CDC leaderboards
- Seasonal runoff forecasts for California’s snow-fed river basins, where the early predictions beat the state’s official Bulletin 120 outlook by a clear margin
- Atmospheric CO2 mapping that fuses GOES-East weather satellite data with other inputs to estimate emissions every 10 minutes, everywhere — picking up both the urban plumes over the Los Angeles basin and plants’ daytime uptake
- 3D solar energy collection tuned with Google Antigravity, in which ERA landed on a 500-triangle volumetric fan that traps scattered radiation with zero backward shading
- Retail forecasting assembled from US economic indicators, Google Trends, and historical patterns, landing at or above Action Economics’ consensus estimates and the Chicago Fed’s CARTS forecasts
The collaborator list runs from epidemiologist Nicholas Reich to climate scientists Tapio Schneider, Ignacio Lopez-Gomez, and Anna Michalak, among others — ERA is positioned to amplify domain experts, not to replace them.
The Gap Between ERA and Ordinary Code Generators
What separates ERA from a general code generator comes down to the objective function. ERA does not chase output that merely “looks like runnable code”; it chases empirical performance on the scientific problem at hand, scored against the success metric the researcher supplies. Structurally, that puts it closer to a virtual postdoc: literature search, method combination, experiment design, and evaluation all sit inside a single loop, and the loop keeps running until the metric improves. A general-purpose coding assistant can turn a prompt into a draft script; ERA hunts through a solution space of thousands of candidate programs and keeps the ones that genuinely do a better job predicting hospital admissions or river runoff.
Google frames the consequence as a matter of access: expert-level computational modeling could be democratized — a real lift for regional research institutes and public-sector teams that lack software engineering capacity.
The Code Is Out; Here Is What Opens Next
The code and experiments behind ERA are open source on GitHub (google-research/era). Up next is Computational Discovery, which pairs ERA with AlphaEvolve and is opening in stages through a trusted tester program under Gemini for Science in Google Labs. It complements two sibling experiments — Hypothesis Generation, created with AI Co-Scientist and published in Nature the very same day, and Literature Insights — with each covering its own stretch of the scientific method. The science page on labs.google is taking registrations from research teams right now.