Google Automates Scientific Discovery With Gemini-Powered ERA

Google researchers led by John Platt have developed Empirical Research Assistance (ERA), an AI system designed to automate scientific discovery by optimizing code for scoreable tasks. The framework, which is open-source on GitHub, functions like an automated Kaggle competitor. It uses a large language model, such as Gemini, to maintain a running tree of past experimental notebooks. Guided by Monte Carlo Tree Search and the Upper Confidence Bound rule, the system optimistically selects promising notebooks and proposes mutations to maximize a target score.

The system's performance saw a massive step change during the transition from Gemini 2.0 to Gemini 2.5, shifting from an ineffective tool to a highly successful one. ERA has already helped Platt's team solve multiple outstanding scientific problems, yielding at least ten published papers. Among these achievements was cracking a counterfactual climate modeling problem regarding how airplane condensation trails reflect sunlight, an obstacle that had previously stumped the Google team for over two years.

This breakthrough is highly relevant to climate science, as contrails account for roughly one percent of human-induced global warming, with just one gram of jet exhaust seeding ten kilograms of ice crystals. For practitioners, ERA shifts the bottleneck of scientific research from tedious code optimization to the conceptual formulation of the scoring metric itself. However, Platt warns that researchers must remain vigilant against reward hacking and overfitting. He pointed to a Google contrail-detection Kaggle competition where the winning entry secured a fifteen-thousand-dollar prize simply by exploiting a half-pixel labeling error rather than solving the underlying physical problem.