Four Ways Google Research's Empirical Research Assistance Is Used by Scientists

Since last fall, Google Research's team has been deploying Empirical Research Assistance (ERA), a tool for generating expert empirical software that helps scientists solve real-world problems. In epidemiology, ERA is used for weekly forecasting of hospitalizations from influenza, COVID-19, and respiratory syncytial virus (RSV) for all US states; Google has ranked among the leaders in public forecasting benchmarks. In cosmology, ERA, combined with Gemini Deep Think, derived six general solutions and a compact formula for the asymptotic limit of gravitational radiation from cosmic strings — a problem previously solved only for a special case. For CO₂ monitoring, scientists used ERA to develop a single-pixel physics-informed neural network that extracts column-averaged carbon dioxide concentrations from data from the geostationary GOES East satellite every 10 minutes, offering significantly higher spatiotemporal resolution compared to the OCO-2 satellite. In neuroscience, ERA, given the zebrafish brain simulator simZFish, proposed interpretable neural mechanisms linking stimulus to motor response, outperforming black boxes in generalization ability. These projects demonstrate ERA's potential for automating scientific discovery across diverse disciplines.