An AI-directed laboratory at Lila Sciences has identified a breakthrough catalyst for green hydrogen production, solving a chemistry problem that has blocked renewable energy progress for 50 years. In a three-month sprint, the autonomous lab proposed, synthesized, and screened 2,942 catalysts, identifying six high-performing material families that could transform how we produce clean hydrogen fuel.

Why Has Green Hydrogen Been So Hard to Make?

Hydrogen is an appealing clean fuel because burning it produces only water as a byproduct. It can power vehicles, generate electricity in fuel cells, or serve as a chemical building block for industrial processes like ammonia production. The vision of a "hydrogen economy" has circulated for decades, but scaling it up has proven nearly impossible.

The bottleneck isn't renewable electricity; it's the chemistry itself. Extracting hydrogen from water requires two chemical reactions. One happens at the cathode, but the other, called the oxygen evolution reaction (OER), happens at the anode in a brutally corrosive acidic environment. This reaction is so harsh that it dissolves most metals. Today's commercial electrolyzers rely on iridium oxide or ruthenium as catalysts, but both are among the rarest elements on Earth. Nearly all iridium is produced as a byproduct of platinum mining, with only a few tonnes produced worldwide annually.

For decades, scientists have searched for a more abundant metal that could survive thousands of hours in acid while remaining active and stable. Nobody had found one, until now.

How Did the AI Lab Make This Discovery?

Lila Sciences, a Cambridge-based AI startup, built an AI Science Factory (AISF) to tackle OER catalyst discovery as its first physical sciences project. The platform combines Bayesian models, which reason well under uncertainty, with language models that have broad context about how chemistry works. This combination allows the AI to ask what it does not yet know, rather than simply optimizing within existing published knowledge.

The workflow operated as a closed loop through four stages: synthesis, pre-test characterization, testing, and post-test characterization. Data from each stage flowed back into the AI model, informing future decisions. The AI analyzed the chemical space and proposed new recipes for materials to synthesize and screen. Initially, electrochemist Ken Jenewein reviewed all suggestions, but he quickly found the AI was landing on picks he would have made himself. Eventually, only safety checks were needed before synthesis began.

The AISF synthesized 96 catalysts in parallel using physical vapor deposition, then quality-checked each one to confirm composition. Each material was then run through accelerated OER screening in acid, measuring both activity and stability.

What Made This Discovery Surprising?

The breakthrough moment came when the AI proposed a metal composition that scientists had long dismissed as unsuitable for acidic OER. When the results appeared on Jenewein's screen, he flagged them as unusual. Senior scientist Fae Habib Zadeh expected the model to still be exploring the chemical space, not converging on strong candidates so quickly.

The team brought the results to John Gregoire, Lila's chief autonomous science officer, and Rafael Gómez-Bombarelli, chief scientific officer of physical sciences at MIT. Between them, they have 40 years of materials science discovery experience. They took one look and decided the model had made an error. The metals listed would not be potent catalysts for OER.

"My former group at Caltech explored catalysts for this reaction for 12 years, and if a student had suggested trying this combination of elements, I would have advised them to try a more promising direction. This combination defies all traditional wisdom," said John Gregoire, chief autonomous science officer at Lila Sciences.

John Gregoire, Chief Autonomous Science Officer at Lila Sciences

They were wrong. After repeated activity tests and over 1,000 hours of stability testing, the top metal composition performed as admirably as ruthenium, one of the industry's standards for OER and one of the rarest metals on the planet.

"We're still scratching our heads. On its own, this well-known catalyst doesn't work for OER. Somehow, the model found the right ingredients to make it active and stable. It is a tangible insight that I'm excited to get in front of the scientific community," explained Rafael Gómez-Bombarelli, chief scientific officer of physical sciences at MIT.

Rafael Gómez-Bombarelli, Chief Scientific Officer of Physical Sciences at MIT

What Does This Mean for Materials Science and AI?

This discovery represents what Lila calls "scientific superintelligence": AI that doesn't just accelerate human research but reaches scientific insights humans wouldn't have discovered on their own. The novel OER catalyst isn't a finished technology yet, and the team is careful to note that more work lies ahead. But it demonstrates what an AI-first, closed-loop laboratory can achieve when engineered for scale and pointed at one of humanity's greatest challenges.

The success hinges on a fundamental shift in how AI approaches discovery. Rather than optimizing within the boundaries of existing knowledge, the system asks what remains unknown. This mirrors the approach that led to breakthroughs in other domains, from chess with Deep Blue to Go with AlphaGo, but applied to the combinatorial challenge of finding new materials.

How AI-Driven Materials Discovery Works in Practice

• Autonomous Synthesis: The AI Science Factory synthesizes 96 candidate materials in parallel using physical vapor deposition, eliminating the bottleneck of one-at-a-time manual synthesis that traditional research faces.
• Real-Time Feedback Loops: Data from characterization and testing stages flows directly back into the AI model within hours, allowing the system to refine its hypotheses and propose new experiments based on actual results rather than theoretical predictions.
• Uncertainty-Driven Exploration: The AI balances what it knows with what it needs to learn, using Bayesian reasoning to identify gaps in its understanding and propose experiments that maximize information gain rather than simply optimizing known parameters.
• Human Oversight at Scale: Scientists remain in the loop for safety and feasibility checks, but the AI calls the shots on experimental design, allowing human expertise to focus on judgment rather than routine decision-making.

The discovery of a high-performing OER catalyst in just three months, after 50 years of human-led research, suggests that AI-directed laboratories may fundamentally reshape how we approach materials science. The challenge now is scaling this approach to other bottleneck problems in energy, manufacturing, and chemistry where the space of possible solutions is too large for traditional experimentation to navigate efficiently.

Lila Sciences publicly launched in March 2025 with the bet that AI could run the full scientific method, hypothesis through iteration, in autonomous laboratories. The OER catalyst discovery suggests that bet is paying off, at least for problems with a large combinatorial space and a clear measure of success.