Four Ways Google Research Scientists Use the Empirical Research Assistance Tool

Since introducing Empirical Research Assistance (ERA) in September, Google Research scientists have used it to address real-world applications beyond proof-of-concept tests. In epidemiology, ERA generated forecasts for influenza, COVID-19, and RSV, competing with CDC tools; Google now submits weekly flu forecasts for all U.S. states and time horizons up to four weeks, and has joined CDC's COVID-19 and RSV forecast hubs, performing at or near the top of public leaderboards. In cosmology, ERA combined with Gemini Deep Think derived six general solutions and a concise asymptotic formula for gravitational radiation from cosmic strings, solving an open problem that previously only had a partial solution for a special case. For atmospheric monitoring, ERA developed a single-pixel physics-guided neural network that extracts column-averaged CO2 from GOES East satellite observations, achieving unprecedented spatial and temporal resolution by combining data from 16 wavelength bands, meteorology, solar angles, and day of year, trained on sparse OCO-2 and OCO-3 data. In neuroscience, ERA used the wiring diagram of a simplified zebrafish brain simulator (simZFish) to propose neural circuits that connect stimulus to motor response, which generalized to new stimuli and offered interpretable, mechanistically accurate solutions, building on earlier results where ERA outperformed baselines in predicting over 70,000 neurons in ZAPBench. These projects showcase ERA's ability to advance science by solving open problems, democratizing computational modeling, and maximizing utility of existing observational data.