MIT’s CRESt laboratory combined AI with automated experiments to discover a new fuel-cell catalyst. Its September 2025 study tested the selected material in a working device.[1]
MIT reported 9.3 times the pure-palladium baseline for power density per dollar of catalyst cost. The cost denominator covers the catalyst, rather than an entire electricity supply system.[2]
Why make electricity from a salt?
Formate offers a way to store energy in a material that is relatively straightforward to handle. Sodium and potassium formate can be held as stable solids, then dissolved in water and fed into a fuel cell. The cell converts chemical energy into electrical energy without burning the fuel.[3]
Earlier MIT and Harvard work demonstrated a small laboratory sequence that captured carbon dioxide, converted it into formate using electricity, and later used the formate to produce power. The attraction is storage: an energy carrier can be manufactured when electricity is available and used at a different time or place.[3]
The catalyst contributes to the final conversion. Its surface must encourage the desired reaction while remaining available for the next fuel molecules. A 2018 electrochemical study of palladium showed how hydrogen attached to the surface can occupy reaction sites and slow formate oxidation. Improving a catalyst therefore involves controlling what sticks to it as well as what reacts.[4]
A material built from eight elements
CRESt’s selected composition combined palladium, platinum, copper, gold, iridium, cerium, niobium and chromium. Its proportions emerged from successive experimental comparisons.[1]
What the CRESt research demonstrated
| Measure | Reported result | Scope |
|---|---|---|
| Experimental campaign | More than 900 chemistries and 3,500 electrochemical tests | Three months of laboratory workSource 1 |
| Cost-specific screening performance | 9.3 times the pure-palladium baseline | Estimated power density per dollar of catalyst costSource 1 |
| Working fuel-cell comparison | Optimised compositions tested against palladium | 60°C; 2 mg of total catalyst per square centimetreSource 1 |
For screening, researchers estimated fuel-cell power from electrochemical measurements using a simplified model. They then tested selected catalysts in working fuel cells. The approximate calculation gave the same relative ranking as the device tests.[1]
The team linked the improvement to the atomic environment of the catalyst’s active sites and its resistance to interfering species. Changing the mixture can help those sites remain useful while reducing dependence on expensive constituents.[2]
Choosing the next experiment
Bayesian optimisation supplies the statistical backbone. It builds a model of how a recipe might perform and how uncertain that prediction is. Choosing an experiment then balances promising candidates against opportunities to learn something useful about an unfamiliar part of the search space.[6]
CRESt additionally uses scientific descriptions and microscope features to guide its search.[1]
The physical system prepares samples, rapidly synthesises materials, examines them and measures electrochemical behaviour. Results feed back into the models that recommend further experiments. A researcher can ask the system to investigate a material recipe or analyse an image through its conversational interface.[2]
The released software shows how practical this integration is. Separate components control liquid handling, electron microscopy and electrochemical testing; a database retains experimental information; and the active-learning code chooses subsequent candidates. Those components connect a proposed recipe to equipment that can make and examine it.[5]
Each completed experiment adds a measured link between composition and behaviour. A promising prediction can be tested, and an unsuccessful recipe can still narrow the next search. That feedback makes a larger experimental campaign useful even when many individual samples do not outperform the current best.
Seeing when an experiment goes wrong
Cameras also help the system investigate inconsistent results. It can notice a shifted sample or a problem during pipetting and suggest explanations. Humans performed most of the debugging in the reported work.[2]
The code is available for other laboratories to examine and adapt. Its equipment modules reveal how the laboratory connected preparation, observation and measurement. Teams with the relevant instruments can reuse those components while adapting the workflow to their own scientific question.[5]
For formate fuel cells, the achievement is a tested improvement in one component that matters economically. For materials research, it is a demonstration of several kinds of evidence contributing to the same experimental decision. Better ingredients and a more productive way of finding them can develop together.
Sources & further reading
- A multimodal robotic platform for multi-element electrocatalyst discovery Nature · 23 Sep 2025 · Accessed 11 Oct 2026
- AI system learns from many types of scientific information and runs experiments to discover new materials Massachusetts Institute of Technology · 25 Sep 2025 · Accessed 11 Oct 2026
- Engineers develop an efficient process to make fuel from carbon dioxide Massachusetts Institute of Technology · 30 Oct 2023 · Accessed 11 Oct 2026
- Adsorbed Hydrogen as a Site-Occupying Species in the Electrocatalytic Oxidation of Formate on Pd/C in Alkaline Medium Journal of The Electrochemical Society / GIST Scholar · Accessed 11 Oct 2026
- CRESt: Automated robotic platform for conducting electrochemical reactions at MIT CRESt researchers / GitHub · Accessed 11 Oct 2026
- BoTorch: A Framework for Efficient Monte-Carlo Bayesian Optimization NeurIPS 2020 · Accessed 11 Oct 2026