Researchers at Google and Harvard have developed Empirical Research Assistance, an Artificial Intelligence system that automatically writes and improves scientific software for specialized research tasks. The system targets what the team calls “empirical software,” custom-built programs designed to maximize performance on a scientific task that can be measured by a numerical score. Such software plays a central role in modern science, where researchers rely on code to test hypotheses, interpret data, and optimize models for problems like weather prediction, disease forecasting, and protein structure prediction.
Empirical Research Assistance is designed to automate the full cycle of scientific software design and refinement, a process that often takes months or even years for human experts. It combines the Google Gemini large language model with a search strategy that explores and refines many possible code variations. Starting from baseline code for a specific problem, the system proposes modifications such as adding components or swapping algorithms, then evaluates whether those changes improve a predefined quality score. It uses tree search, a method also used in systems like AlphaGo, to decide which approaches to pursue and which to discard.
The system can also incorporate research ideas from papers and textbooks, either supplied directly by a user or retrieved automatically, and fold those ideas into later versions of the code. That design allows it to recombine existing concepts in ways that may uncover promising solutions that researchers would be unlikely to test manually. The work was co-led by Michael Brenner, Catalyst Professor of Applied Mathematics and Physics at the Harvard John A. Paulson School of Engineering and Applied Sciences and a Google research scientist, along with Shibl Mourad from Google DeepMind. Harvard Ph.D. students Qian-Ze Zhu, Ryan Krueger, and Sarah Martinson contributed as Google student researchers while working in Brenner’s group.
In testing, the system was applied to several scientific problems. Zhu used it to predict the activity of more than 70,000 neurons in the brain of a zebrafish and compare the results against actual neural data. In one experiment, the team prompted Empirical Research Assistance to use an existing neuron-modeling library to build more physically accurate simulations of neural activity, a task that would otherwise have required substantial manual effort to learn and tune. Zhu said methods that previously might take a week to implement can now be run in parallel in a few hours.
On one test, the ERA system generated 14 models for predicting COVID-19 hospitalizations that outperformed the best U.S. Centers for Disease Control models used during the pandemic. In another experiment, ERA discovered four new methods for integrating single-cell RNA sequencing datasets, beating top human-designed approaches. The researchers say the system could cut exploration time from months to hours or days, potentially freeing scientists to focus more on defining important questions and tackling creative and critical research challenges.
