AI System Automates Coding for Scientific Research

Google and Harvard researchers developed an AI system called Empirical Research Assistance (ERA) that automates the creation of high-performance scientific software, surpassing human-written programs and accelerating research across fields like chemistry and neuroscience. The system combines Google’s Gemini language model with tree search techniques to refine code, integrate research ideas, and solve complex tasks such as predicting disease spread or protein shapes, as demonstrated in a study published in *Nature*." "article": "A team of researchers from Google and Harvard has created an AI system called Empirical Research Assistance (ERA) that automatically generates scientific software programs outperforming those written by humans. The project, co-led by Michael Brenner, Catalyst Professor of Applied Mathematics and Physics at Harvard’s John A. Paulson School of Engineering and Applied Sciences, and Shibl Mourad from Google DeepMind, was published in *Nature*. ERA automates the design and refinement of empirical software—custom programs optimized for specific scientific tasks like weather prediction or disease modeling. Such software underpins breakthroughs in fields such as chemistry, but developing it manually is time-consuming. ERA eliminates this bottleneck by using Google’s Gemini language model to explore and refine thousands of code variations faster than human experts. The system employs tree search, a method also used in AI like AlphaGo, to evaluate and refine code modifications. It improves predefined quality scores, such as predicting hospitalization rates or protein shapes, by proposing algorithmic changes. ERA can also integrate research ideas from papers or textbooks, either provided by users or retrieved automatically, allowing it to uncover solutions humans might overlook. The team tested ERA on diverse problems, including predicting neuron activity in zebrafish brains. The system’s ability to combine research ideas efficiently makes it a powerful tool for scientific discovery. Brenner noted that ERA’s integration of research concepts enables it to find solutions that would otherwise remain untested. ERA represents a significant advancement in AI-driven scientific research, potentially accelerating discoveries across multiple disciplines.