Google Unveils ERA for Scientific Coding and Discovery
Google has published a paper in Nature describing Google ERA (Empirical Research Assistance), an AI system that writes and optimizes scientific code by evaluating thousands of possible solutions through a tree-search method. Alongside the publication, Google announced that ERA now underpins a new experimental tool called Computational Discovery, which is beginning to roll out through Gemini for Science - making the leap from research paper to accessible product.
How ERA Works and What the Nature Paper Shows
ERA uses Google's Gemini model to tackle one of the slowest parts of research: writing, testing, and refining computational experiments. Given a scientific problem and a success metric, the system searches relevant literature, generates code, explores multiple solution paths, and combines techniques - all while scoring its output against the stated goal.
According to Google's research blog, the Nature paper tested ERA across six benchmark domains: genomics, public health, satellite imagery analysis, neuroscience prediction, time-series forecasting, and mathematics. The system achieved what Google describes as "expert-level performance" across all of them, suggesting it could broaden access to sophisticated computational modeling for researchers who lack deep coding expertise.
Real-World Applications Already Underway
Google Research scientists have spent the past six months applying ERA to open scientific questions. The team has now released eight manuscripts covering specific use cases, five of them new.
Among the highlights: ERA-built models for predicting U.S. hospital admissions from flu, COVID-19, and RSV consistently ranked at or near the top of CDC forecasting leaderboards, according to Google. In another project, ERA produced seasonal water-runoff forecasts for California's snowfed river basins that were significantly more accurate than the state's official Bulletin 120 outlook.
The system was also used to map atmospheric CO2 concentration with what Google calls "unprecedented spatial and temporal resolution," using geostationary satellite data. Other projects explored 3D solar panel optimization - combining ERA with Google Antigravity - and macroeconomic retail forecasting using public economic indicators and Google Trends data.
From Research to Google Labs
The bigger story may be what happens next. Google announced that Computational Discovery, built with both ERA and AlphaEvolve, is now gradually opening access through labs.google/science. It sits alongside two other new Gemini for Science experiments: Hypothesis Generation (powered by AI Co-Scientist, also published in Nature today) and Literature Insights.
Google frames these three tools as complementary, each supporting a different stage of the scientific method. Computational Discovery handles the code-heavy experimentation phase, Hypothesis Generation aids early-stage ideation, and Literature Insights helps researchers navigate existing work.
The transition from a peer-reviewed paper to a live experimental tool is worth watching. Whether ERA can deliver expert-level results in the hands of a broader scientific community - outside Google's own benchmarks - will be the real test as access expands.