(Krisana Antharith/Shutterstock)
AI-powered tools have become more common in scientific research and development, especially for predicting outcomes or suggesting possible experiments using datasets. However, most of these systems only work with limited types of data. They might rely on numbers from a few tests or chemical inputs, but that only scratches the surface.
Human scientists bring much more to the table. In a lab, decisions are shaped by a mix of sources. Researchers consider published papers, past results, chemical behavior, images, personal judgment, and feedback from colleagues. That kind of depth is hard to replace. No single piece of information tells the whole story, and it is the combination that often leads to real breakthroughs. However, humans can’t match the sheer processing ability of AI systems.
A new platform developed at MIT, named Copilot for Real-world Experimental Scientists (CRESt) is designed to work more like a true research partner. The system pulls together many kinds of scientific information and uses that input to plan and carry out its own experiments.
CRESt builds on active learning but expands beyond it by using multimodal data. It learns from what it sees, adapts based on results, and continues to improve over time. For fields like materials science, where progress often takes years, CRESt offers a faster and more complete way to search for new ideas.
“In the field of AI for science, the key is designing new experiments,” says Ju Li, School of Engineering Carl Richard Soderberg Professor of Power Engineering. “We use multimodal feedback — for example information from previous literature on how palladium behaved in fuel cells at this temperature, and human feedback — to complement experimental data and design new experiments. We also use robots to synthesize and characterize the material’s structure and to test performance.”
(nevodka/Shutterstock)
The researchers behind CRESt wanted to create something that felt less like a computer program and more like a working partner in the lab using data. They aimed to build a system that could follow the full rhythm of experimental science, not just react to isolated bits of data.
The full study describing CRESt and its results was published in Nature. A key aim with CRESt is to enable scientists to speak to it naturally using AI. For example, they can get help with tasks like reviewing microscope images, testing new material combinations, or making sense of earlier results. Once a request is made, the system searches through what it knows, sets up the experiment, runs it through automated tools, and uses the outcome to shape what comes next. The process keeps going, with each round of testing feeding into the next stage of learning.
Reproducibility has long been a challenge in labs, but the team explained that CRESt helps by watching experiments as they happen. With cameras and vision-language models, it can flag small errors and suggest fixes. The researchers said this led to more consistent results and greater confidence in their data.
The team said that basic Bayesian optimization was too narrow, often stuck adjusting known elements. CRESt avoids that limit by combining data from literature, images, and experiments, then exploring beyond a small box of options. This broader reach was critical in its fuel cell work.
The research team chose fuel cells as one of the first areas to test CRESt, a field where progress has often been slowed by the size of the search space and the limits of conventional experimentation. According to the team, the system combined information from published papers, chemical compositions, and structural images with fresh electrochemical data from its own tests. Each cycle added more results to its dataset, which was then used to refine the next set of experiments.
In three months, CRESt evaluated more than 900 different chemistries and carried out 3,500 electrochemical trials. The researchers report that this process led to a multielement catalyst that relied on less palladium but still delivered record performance.
(amgun/Shutterstock)
“A significant challenge for fuel-cell catalysts is the use of precious metal,” says Zhang. “For fuel cells, researchers have used various precious metals like palladium and platinum. We used a multielement catalyst that also incorporates many other cheap elements to create the optimal coordination environment for catalytic activity and resistance to poisoning species such as carbon monoxide and adsorbed hydrogen atom. People have been searching low-cost options for many years. This system greatly accelerated our search for these catalysts.”
According to the team, CRESt was not built to simply run one experiment after another. Before a test is carried out, the system reviews information from past studies, databases, and earlier results to build a picture of what each recipe might mean. That broader view helps narrow the field of options so the experiments that follow are more focused.
Each new round of testing adds to the record, and those results, combined with feedback from researchers, are folded back into the system. The researchers shared that this cycle of preparation, testing, and refinement was central to the speed with which CRESt was able to move through hundreds of possible chemistries during the fuel cell work.
The researchers emphasize that CRESt is not designed to replace scientists. “CREST is an assistant, not a replacement, for human researchers,” Li says. “Human researchers are still indispensable. In fact, we use natural language so the system can explain what it is doing and present observations and hypotheses. But this is a step toward more flexible, self-driving labs.” With impressive initial results, it appears MIT might have developed a platform that gives scientists a new kind of partner in the lab.
Related Items
How Scientists Are Teaching AI to Understand Materials Data
Argonne: Turning Materials Data into AI-Powered Lab Assistants
MIT’s CHEFSI Brings Together AI, HPC, And Materials Data For Advanced Simulations
