Google has published research on Empirical Research Assistance (ERA) in Nature — a tool that uses Gemini to write and optimize scientific code. ERA tackles one of research's most time-consuming tasks: iteratively testing and refining computational experiments. The company is now making this technology accessible through Computational Discovery, rolling out via Gemini for Science.

Using a tree search approach, ERA evaluates thousands of options to optimize output code against given scientific goals. Testing across benchmarks in genomics, public health, satellite imagery, neuroscience, time-series forecasting, and mathematics showed expert-level performance across all domains. This could democratize access to advanced computational modeling while expanding what experienced researchers can accomplish.

Real-world applications demonstrate the tool's capabilities. ERA-driven epidemiological forecasts consistently rank at or near the top of CDC leaderboards for flu, COVID-19, and RSV. California water managers are using it for more accurate spring runoff predictions than the state's official Bulletin 120 outlook. The system also mapped atmospheric CO2 from satellite data and optimized solar panel designs with a 500-triangle volumetric fan configuration.

Google is now expanding access to Computational Discovery, which combines ERA with AlphaEvolve, through labs.google/science. The launch includes Hypothesis Generation built with AI Co-Scientist, also published in Nature today. Eight manuscripts now apply ERA to specific scientific problems.