Empirical Research Assistance (ERA): From Nature publication to catalyzing Computational Discovery

Google Research just published the kind of paper that tends to get buried under bigger AI headlines, but it might matter more than most. ERA, short for Empirical Research Assistance, is a Gemini-powered system built to do one specific, grinding job: write and refine the code scientists use to run experiments. Anyone who's spent a week tweaking a model's hyperparameters knows why that's worth automating. ERA runs a tree search over thousands of code variants, testing and rewriting until it hits a target metric, and according to the Nature paper it now performs at expert level across genomics, public health, satellite imagery, neuroscience, time-series forecasting and math benchmarks.

What's more interesting than the benchmark scores is what happened after Google let researchers actually use it. Over six months, teams applied ERA to real open problems and the results are now piling up as papers, eight so far. One team used it to forecast U.S. hospital admissions for flu, COVID and RSV, and those forecasts have been sitting at or near the top of the CDC's public leaderboards. Another pointed ERA at California's snowmelt-fed rivers and produced spring runoff predictions that beat Bulletin 120, the state's own official water outlook — a genuinely useful edge for a state that fights over water allocation every summer.

The carbon dioxide work is the one that sticks with me. Using data from a geostationary weather satellite, ERA built a model with enough spatial and temporal resolution to see cities exhaling CO2 and crops inhaling it during daylight hours. That's not a toy result; it's the kind of granular atmospheric picture that climate scientists have wanted for years and haven't had the tooling to build cheaply. Separately, a solar-energy experiment paired ERA with Google's Antigravity system and landed on a 500-triangle fan-shaped panel geometry that traps scattered light with no backward shading — an oddly specific design nobody would have hand-engineered.

Google is now rolling this out beyond its own labs as Computational Discovery, part of a broader Gemini for Science push that also includes a hypothesis-generation tool built on AI Co-Scientist and a literature-review assistant. The pitch is that ERA and AlphaEvolve together can hand working scientists — not just Google's own researchers — a system that searches literature, writes code and iterates toward a result, compressing months of trial and error. Whether outside labs get meaningful access, or just a demo wrapped in a waitlist, is the part worth watching over the next year.