A robotic laboratory steered by an interpretable AI has found new hydrogen-producing catalysts made from earth-abundant metals that, its builders say, are “at least comparable to Pt-based ones”. Chemists at UCLA and National Taiwan University, working with data scientists at the University of North Carolina at Chapel Hill, reported the lab-scale result in Nature Communications on 23 October.
The catalysts drive the hydrogen evolution reaction, the half of water splitting that releases hydrogen gas. Platinum is the reference material for that job. The team’s system tested about 1,200 recipes in roughly four months, work the authors estimate would have taken human researchers at least 600 days.
The robot searched a 12-dimensional recipe space
Each recipe is an electroplating bath. The platform mixes precursors for seven metals (vanadium, chromium, iron, cobalt, nickel, copper, and magnesium as a deliberate dud), plus sulfur, selenium and phosphorus, then sets the plating voltage and duration. That makes 12 variables, far too many to search by hand or by a simple grid.
According to the paper, the automated platform plates a material onto a carbon electrode, measures its activity in a dilute alkaline solution (0.1 M potassium hydroxide), strips the electrode clean and moves to the next recipe. The supplementary information puts the cycle at roughly 2.5 hours per material, about ten a day, against about two a day for a student working manually. Human input was limited to around 30 minutes every two days to top up solvents.
The authors describe the hardware as “relatively affordable”, at $27,050 in 2025 excluding the potentiostat, the instrument that controls and measures the electrochemistry. Chong Liu’s group at UCLA says on its publications page that the work builds on its earlier projects in automated and machine-learning electrochemistry.
A chemist’s hunch is built into the algorithm
The decision-making software is what the team says sets the work apart. Most self-driving labs use what the authors call “black-box models that are difficult to interpret”. This one uses a bandit algorithm, a type of reinforcement learning that chooses each next experiment, and its maths is shaped by chemistry.
Catalyst activity often follows a “volcano” curve: bind hydrogen too weakly or too strongly and performance falls. The team assumed activity should therefore behave roughly like a quadratic surface close to the best recipes, and built that assumption into what it calls a locally quadratic bandit. Fitted to a training set of about 600 recipes, the model scored the highest R² (up to 0.70) of 20 machine-learning models tested.
Five campaigns followed. Each ran 60 broad searches and then 60 local ones around the most promising recipe. In simulations, the paper reports, the approach converged “noticeably faster” than a standard black-box method, Bayesian optimisation.
The winners came from iron-cobalt and copper-nickel chemistries
Researchers then remade the best recipes by hand and tested them on standard lab electrodes. One, mostly partially oxidised iron and cobalt phosphides doped with chromium and vanadium, reached an overpotential (the extra voltage needed to drive the reaction) of 0.039 V at a low current density of 2 mA per square centimetre. Another, mostly metallic copper and nickel phosphide, reached 0.198 V at 50 mA per square centimetre.
The team compared these with a platinum electrode and published values, calling them “competitive relative to Pt and representative alkaline HER benchmarks under similar conditions”. The paper reports no long-running stability tests, and nothing was built at device scale; it mentions only “the benefits of potential scalability” for electroplated materials.
Because the model is quadratic, it also produces a map of how pairs of ingredients interact. For the iron-cobalt catalyst, the ratios among cobalt, iron and phosphorus mattered most, while magnesium’s influence shrank as the run went on, as expected for an inactive control. The authors say these maps are descriptors of the model, “not as the direct experimental proof of catalytic mechanisms”.
The interpretable model is a bigger result than any one catalyst
Plenty of robotic labs can find a good material. The more useful claim here is that this one explains its choices in terms a chemist can test. A quadratic model is deliberately simple, and an R² of 0.70 is modest, but simplicity is the point: a ranked list of ingredient pairs gives the next experiment a hypothesis rather than a black-box score. It is a different test for AI in science from AI-driven weather forecasting, where accuracy is the goal: here the measure is whether the humans learn something.
The catalyst figures need context. The lowest overpotential was measured at a gentle current density, on small lab electrodes, in a dilute alkaline solution. A few lab measurements say little about commercial use, and the paper offers no data on how these materials age. “Comparable to platinum” is the authors’ own verdict on lab measurements, not a finding about working hardware.
It belongs with the recyclable flow battery and the ultra-thin carbon membrane published this week: clean-energy chemistry that is promising in the lab and unproven at scale. Our weekly tech for good roundup tracks how many of these results go further.
The authors say the method “is generally applicable to any chemistry problems”, provided researchers can write their knowledge into the model. The question the paper leaves open is the one industry will ask first: how long an iron-cobalt or copper-nickel electrode keeps working.
Sources
- Lai, Y.-A. et al., “Chemistry-infused bandit framework for interpretable autonomous campaign of electrocatalysis”, Nature Communications, 23 October 2026
- Supplementary Information for the paper (PDF)
- Liu Lab, UCLA: Publications
Related stories
- Recyclable Flow Battery Kept 98.1% of Capacity Over 2,200 Lab Cycles
- Tech for Good This Week: AI Diagnosis, Cleaner Grids and Scam Busts
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
