A new artificial intelligence system can automatically write high-performance scientific software, in some cases outperforming programs built by human experts, according to a study published in Nature. The team behind it says the tool could meaningfully speed scientific discovery across many disciplines.
The system, called Empirical Research Assistance (ERA), was developed by a team at Google co-led by Michael Brenner from Harvard University and a Google research scientist, along with Shibl Mourad of Google DeepMind.
Across modern science, custom-built software is routinely used to test hypotheses or interpret complex data. The authors describe this category as "empirical software"—code whose sole purpose is to maximize performance on a scientific task such as predicting weather or forecasting hospitalizations during a disease outbreak. Any such problem framed as a numerical score is what they call a scorable task. Empirical software underpins major scientific advances, including three recent chemistry Nobel prizes, but designing and refining it can demand months or years of expert work.
ERA was designed to remove that bottleneck by automating the entire design-and-refinement cycle. It pairs Google's Gemini large language model with a tree-search strategy similar to those used in game-playing systems like AlphaGo. Starting from a baseline piece of code, the system generates and tests thousands of modifications—adding components, swapping algorithms, and drawing on ideas pulled from research papers or textbooks—to improve a predefined quality score.
"This ability to integrate and recombine research ideas enables the system to find needle-in-a-haystack solutions that human research might never get to test," Brenner said. The team applied ERA to a range of scientific problems. In one test, it was used to predict the activity of more than 70,000 neurons in the brain of a zebrafish and compare the predictions with actual neural data. A task that would have taken weeks or months of learning a new software package was handled automatically.
"This new system is going to accelerate scientific discovery by allowing you to explore a lot of ideas at the same time," co-athor Qian-Ze Zhu said. "Previously it might take you a week to implement some specific methods, but now you can just run them in parallel in a few hours." ERA also produced 14 COVID-19 hospitalization models that outperformed the best U.S. Centers for Disease Control models used during the pandemic, and it discovered four new methods for integrating single-cell RNA sequencing datasets that beat top human-designed approaches.
The system, called Empirical Research Assistance (ERA), was developed by a team at Google co-led by Michael Brenner from Harvard University and a Google research scientist, along with Shibl Mourad of Google DeepMind.
Across modern science, custom-built software is routinely used to test hypotheses or interpret complex data. The authors describe this category as "empirical software"—code whose sole purpose is to maximize performance on a scientific task such as predicting weather or forecasting hospitalizations during a disease outbreak. Any such problem framed as a numerical score is what they call a scorable task. Empirical software underpins major scientific advances, including three recent chemistry Nobel prizes, but designing and refining it can demand months or years of expert work.
ERA was designed to remove that bottleneck by automating the entire design-and-refinement cycle. It pairs Google's Gemini large language model with a tree-search strategy similar to those used in game-playing systems like AlphaGo. Starting from a baseline piece of code, the system generates and tests thousands of modifications—adding components, swapping algorithms, and drawing on ideas pulled from research papers or textbooks—to improve a predefined quality score.
"This ability to integrate and recombine research ideas enables the system to find needle-in-a-haystack solutions that human research might never get to test," Brenner said. The team applied ERA to a range of scientific problems. In one test, it was used to predict the activity of more than 70,000 neurons in the brain of a zebrafish and compare the predictions with actual neural data. A task that would have taken weeks or months of learning a new software package was handled automatically.
"This new system is going to accelerate scientific discovery by allowing you to explore a lot of ideas at the same time," co-athor Qian-Ze Zhu said. "Previously it might take you a week to implement some specific methods, but now you can just run them in parallel in a few hours." ERA also produced 14 COVID-19 hospitalization models that outperformed the best U.S. Centers for Disease Control models used during the pandemic, and it discovered four new methods for integrating single-cell RNA sequencing datasets that beat top human-designed approaches.
