Empirical Research Assistance: Accelerating Scientific Discovery with AI-Driven Coding

The speed and scope of scientific discovery are often bottlenecked by the tedious, iterative process of writing, testing, and refining computational experiments. To address this, Google has introduced Empirical Research Assistance (ERA), a system powered by Gemini that automates the optimization of scientific code. Recently detailed in a publication in Nature, ERA represents a shift toward "Computational Discovery," where AI doesn't just assist in writing code, but actively optimizes it to achieve specific scientific goals.

How ERA Works: Beyond Simple Code Generation

Unlike standard AI coding assistants that provide a single suggestion based on a prompt, ERA is designed for the empirical nature of scientific research. Given a scientific problem and a defined measure of success, ERA performs the following steps:

1. Literature Search: It scans scientific literature to identify relevant methods and techniques.
2. Code Synthesis: It writes the initial computational code to address the problem.
3. Iterative Exploration: Using a tree search approach, ERA considers thousands of potential options, combining different techniques and exploring various solutions.
4. Evaluation and Optimization: It continuously evaluates the results against the goal, refining the code until it reaches an optimal state.

This loop allows ERA to move beyond simple pattern matching, enabling it to discover expert-level computational models that might have been overlooked by human researchers.

Expert-Level Performance Across Disciplines

To validate ERA's capabilities, Google tested the system across a diverse array of benchmarks, including genomics, public health, satellite imagery analysis, neuroscience prediction, time-series forecasting, and mathematics. The results indicate that ERA achieves expert-level performance across all these domains, suggesting a democratization of high-end computational modeling for scientists who may not be expert programmers.

Real-World Scientific Applications

Google has applied ERA to several open scientific questions, producing a series of manuscripts that demonstrate its immediate impact:

Public Health and Environment

• Epidemiological Forecasting: ERA was used to predict U.S. hospital admissions for flu, COVID-19, and RSV. The resulting forecasts consistently ranked at or near the top of CDC leaderboards.
• California Water Management: ERA developed a forecasting model for seasonal runoff in California's snow-fed river basins. This model outperformed the state's official seasonal water supply outlook (Bulletin 120), providing more accurate early predictions for a critical resource.
• Atmospheric CO2 Mapping: By utilizing geostationary weather satellite data, ERA created a model to map CO2 concentrations with unprecedented resolution, capturing human-driven urban enhancements and the natural absorption cycles of plants.

Engineering and Economics

• Solar Energy Optimization: In collaboration with Google Antigravity, ERA optimized solar panel topographies. It discovered a "500-triangle volumetric fan" design that maximizes energy capture by trapping scattered radiation without backward shading.
• Retail Forecasting: ERA applied its capabilities to macroeconomic retail sales forecasting, meeting or exceeding the accuracy of the Chicago Fed's CARTS monthly retail forecast and other commercial consensus estimates.

The Path Toward Computational Discovery

ERA is now a foundational component of a broader suite of tools under the Gemini for Science initiative. It works alongside other experimental tools such as:

• Computational Discovery: A tool built with ERA and AlphaEvolve to automate the discovery of computational models.
• Hypothesis Generation: Powered by AI Co-Scientist, this tool assists in the earlier stages of the scientific method by proposing new hypotheses.
• Literature Insights: A tool designed to streamline the synthesis of existing scientific knowledge.

By integrating these tools, Google aims to support different stages of the scientific method—from reading the literature and generating hypotheses to the writing and optimization of the empirical code required to prove them.