New AI system automatically writes high-performance scientific software

2026-05-21
2 min read.
The ERA system developed by Google and Harvard researchers dramatically speeds up the creation of custom programs used in scientific research and outperforms human-written code.
New AI system automatically writes high-performance scientific software
Credit: Tesfu Assefa

Researchers from Google and Harvard have developed a new artificial intelligence (AI) system that can automatically write scientific software programs. These programs perform better than those created by human experts. The system is called empirical research assistance, or ERA. It is described in a paper published in Nature.

In modern science, researchers rely on special computer programs known as empirical software. These programs are built solely to solve specific scientific problems that can be judged by a numerical score. Such problems are called scorable tasks. Examples include predicting the spread of a disease from past data or determining the three-dimensional shape of a protein from its amino acid sequence. Normally, creating and improving these programs requires experts to test and refine the code many times, a process that can take months or even years.

The era system removes this bottleneck by automating the full cycle of code design, testing, and improvement. It combines a large language model with a search method called tree search. Tree search is a step-by-step technique that explores thousands of possible code changes, similar to how some game-playing artificial intelligence systems evaluate future moves.

How the system performs on real scientific problems

The era system can also take research ideas from scientific papers or textbooks and combine them into the code. This ability lets it discover useful solutions that human researchers might never test. In experiments, ERA created 14 models for predicting covid-19 hospitalizations that worked better than the best models used by the U.S. Centers for Disease Control during the pandemic. In another test, it found four new methods for combining single-cell RNA sequencing data sets, outperforming the best human-designed approaches. It also built accurate models of brain activity in zebrafish far more quickly than a human could have done.

By reducing the time needed to explore ideas from months to hours or days, era frees scientists to focus on truly creative and important questions.

#AIApplications

#LargeLanguageModels(LLMs)



Related Articles


Comments on this article

Before posting or replying to a comment, please review it carefully to avoid any errors. Reason: you are not able to edit or delete your comment on Mindplex, because every interaction is tied to our reputation system. Thanks!