MIT researchers have unveiled a powerful new approach to materials science: letting artificial intelligence guide laboratory robots toward breakthrough discoveries. The result: a fuel cell catalyst that outperforms industry standards by 9.3 times in power density per dollar, accomplished in just 90 days.
The discovery comes from CRESt, the Copilot for Real-world Experimental Scientists, a platform that fuses large language models, automated wet-lab robotics, and Bayesian optimization into a closed-loop system for accelerated materials discovery. Published in Nature, the research demonstrates that AI-human collaboration can solve hard problems in the clean energy transition that have stalled for decades.
The Challenge: Precious Metals and Power Density
Fuel cells promise clean energy for heavy-duty vehicles and industrial applications, but a critical barrier remains: cost. Today's most effective fuel cell catalysts rely on palladium and platinum, rare, expensive precious metals that limit commercial deployment. Researchers have long sought alternative compositions that could maintain performance while reducing precious metal loading.
Traditional materials discovery proceeds by trial and error, with human chemists manually formulating hypotheses, running experiments, and iterating. The bottleneck is obvious: a single experiment takes time, cost, and human expertise. Scaling discovery to thousands of candidate materials remains impractical under conventional workflows.
How CRESt Works: The AI-Robot Loop
CRESt reimagines this workflow by automating the hypothesis-experiment-learning cycle. The system begins by ingesting diverse data sources: prior literature on how palladium behaves in fuel cells, databases of existing materials, and chemical knowledge encoded in large language models. This foundational knowledge is translated into an initial hypothesis about promising catalyst compositions.
Robotic systems then synthesize candidate materials and run electrochemical tests at scale. A liquid-handling robot prepares samples. A carbothermal shock system rapidly synthesizes materials. An automated electrochemical workstation measures performance. Automated electron microscopy and optical microscopy characterize the resulting structures. All observations feed back into multimodal AI models that refine predictions and suggest the next round of experiments.
Unlike pure computational approaches that predict materials in silico, CRESt validates predictions in real experiments. This grounding in physical reality accelerates learning: the system sees what actually works, not just what theory predicts should work.
The Discovery: Eight Elements, Record Power
After exploring more than 900 chemistries and conducting 3,500 electrochemical tests over three months, CRESt identified an eight-element catalyst composition optimized for formate fuel cells. The results are striking: 9.3 times higher power density per dollar compared to pure palladium. In working fuel cells, the CRESt-discovered catalyst delivered record power density while using only one-fourth the precious metals of previous state-of-the-art devices.
As Zhen Zhang, a member of the research team, noted: a significant challenge for fuel-cell catalysts is the use of precious metal, and the team used a multielement catalyst that also incorporates many other cheap elements to create the optimal coordination environment for catalytic activity.
“We use multimodal feedback from previous literature and human feedback to complement experimental data and design new experiments”— Ju Li, Carl Richard Soderberg Professor of Power Engineering, MIT
The dramatic precious-metal reduction, 75 percent less palladium per cell, has immediate implications for manufacturing cost and supply-chain resilience.
CRESt in Context: A Growing AI Materials Ecosystem
CRESt is not alone in the race to accelerate materials discovery through AI. Google DeepMind's GNoME predicted 2.2 million new stable crystal structures by training graph neural networks on existing materials databases. Microsoft's MatterGen platform generates materials with targeted properties, recently discovering a new solid-state battery material in partnership with the Pacific Northwest National Laboratory. Berkeley Lab's autonomous A-Lab facility uses robots guided by AI to synthesize novel materials in a fully automated setting.
What distinguishes CRESt is its focus on solving a specific, high-impact problem, fuel cell catalysis, through a tightly integrated loop of LLM reasoning, physical experimentation, and Bayesian optimization. While competitors emphasize breadth (generating millions of candidates), MIT's approach emphasizes depth: fewer candidates explored more rigorously, with every step grounded in physical validation.
Implications for Clean Energy and US Manufacturing
The fuel cell catalyst market represents a crucial node in the hydrogen economy and green energy transition. Breakthroughs in performance and cost directly enable deployment of fuel cell vehicles and industrial power systems. A 9.3-fold improvement in power density per dollar could compress adoption timelines for heavy-duty zero-emission vehicles, a priority for US climate policy.
Beyond the specific catalyst, CRESt demonstrates a model for US materials manufacturing competitiveness. Rather than relying solely on computational prediction, the platform validates discoveries in real experimental workflows, a requirement for scaling to commercial production. The integration of AI with robotic infrastructure echoes strategic priorities articulated in the National Science Foundation and Department of Energy funding initiatives.
Professor Ju Li, who led the research, emphasized the complementary roles of AI and human expertise: CRESt is an assistant, not a replacement, for human researchers. Human researchers are still indispensable.
The Path Forward
The CRESt platform is now being adapted for other materials challenges: battery electrodes, water electrolyzers, and photocatalysts for CO2 conversion. Each application could benefit from the same closed-loop integration of AI reasoning and robotic validation. As the technology matures, we can expect similar breakthroughs in other domains where materials properties directly constrain energy or environmental outcomes.
The broader message is simple: AI-guided experimental systems compress discovery cycles. What once took years can now take months. In the race to decarbonize, compressed timelines matter.