AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution

Making hydrogen fuel from water requires splitting the water molecule into its parts, a process that demands a great deal of energy to push the reaction forward. One half of this reaction, where oxygen is released, is notoriously difficult to speed up, especially when the process happens in a strong acid. For decades, the only materials reliable enough to handle this harsh, acidic environment without falling apart or losing their ability to work are oxides of iridium and ruthenium. These metals are incredibly rare, and their supply chains are fragile, concentrated in a few places and vulnerable to disruption. If the world hopes to build hydrogen energy systems on a massive scale, relying on such scarce materials creates a bottleneck that could stall the entire industry. Scientists have long searched for a substitute that is both abundant and durable, but finding a material that can survive the acidic conditions while remaining efficient has proven to be an elusive challenge.

A team of researchers at Lila Sciences has now used a new kind of automated discovery platform to find a promising solution, identifying a family of catalysts based on palladium that could serve as a viable alternative. Instead of relying on human intuition to guess which chemical combinations might work, they built a closed-loop system where artificial intelligence directs the experiments. This platform combines high-speed robotic synthesis with machine learning models that learn from every result. The system creates thousands of tiny samples of different metal oxides, tests them for how well they produce oxygen and how long they last in acid, and then uses that data to decide which new samples to make next. The process is largely automated, with robots handling the synthesis and testing while the AI analyzes the results and proposes the next round of candidates, creating a rapid cycle of discovery that moves far faster than traditional methods.

Over the course of their campaign, the system evaluated nearly three thousand different catalyst compositions, exploring a vast landscape of materials made from twenty-six different elements. The goal was to find a balance between two competing needs: high activity, which means the catalyst works with minimal wasted energy, and high stability, which means it does not dissolve or degrade over time. Most of the materials tested failed on one count or the other, either working well but breaking down quickly, or lasting a long time but requiring too much energy to function. However, the AI successfully navigated this difficult trade-off and identified two specific mixtures that stood out. These catalysts are primarily made of palladium, a metal that is far more common than iridium, but they are mixed with tiny amounts of other elements to stabilize them. One of the leading candidates, a mixture of palladium with indium and manganese, proved to be exceptionally durable.

When the researchers put the best candidates through long-term testing in a strong acid solution, the results were striking. A standard palladium oxide catalyst began to degrade after about two hundred hours of operation, while a ruthenium benchmark, the current industry standard, showed a rapid increase in overpotential at 615 hours coincident with the delamination of the film. The new palladium-based catalyst mixed with nickel and tantalum lasted nearly twice as long as the standard palladium version, operating for about four hundred seventy hours before showing significant decline. The champion of the group, the indium-manganese-palladium mixture, maintained its performance for over one thousand hours without failing. During this extended period, it kept the energy required to produce oxygen low, staying well below the threshold where the material would typically begin to break down. This endurance was not just a matter of the material being tough; the researchers found that the catalyst physically changed its shape during operation. Under the microscope, the surface of the successful catalyst developed a unique, needle-like nanostructure that formed naturally as it worked. This structure seemed to protect the underlying metal from corrosion while keeping the surface active for the chemical reaction.

The discovery challenges previous assumptions in the field. For a long time, scientists believed that pure palladium oxide was not a good candidate for this specific job because it was thought to be too inactive or unstable in acid. The conventional wisdom suggested that researchers should look elsewhere, perhaps toward other metals entirely. The AI system, however, did not rely on these old rules. By exploring the data without human bias, it found that adding small amounts of specific elements to palladium could unlock a level of stability and activity that no one had predicted. The system also proved that a flexible, adaptive approach to discovery is superior to rigid, pre-planned methods. When the researchers compared their AI-driven strategy against fixed algorithms and even against large language models that relied on existing scientific knowledge, the adaptive system found the best solutions faster and discovered the new palladium family in fewer attempts. The fixed algorithms tended to get stuck exploring the same known areas, while the language models, lacking the ability to learn from the specific experimental data in real time, failed to identify the promising new materials.

This work does not claim to have solved the problem of hydrogen production entirely, but it offers a significant step forward. Palladium is still a precious metal, but it is much more available than iridium, with a supply chain that is more diversified and less prone to sudden shortages. The ability to use palladium-based catalysts could ease the pressure on global resources while the search continues for even more abundant, non-precious metal alternatives. The researchers emphasize that the true breakthrough lies not just in the specific materials they found, but in the method used to find them. By integrating artificial intelligence with high-throughput experimentation, they created a system capable of uncovering non-obvious solutions that human experts might overlook. The platform successfully demonstrated that machines can learn to navigate the complex trade-offs of materials science, identifying durable, efficient catalysts that can withstand the harsh conditions required for the future of clean energy. The needle-like structures that formed on the surface of the best catalysts suggest that the path to better materials may involve designing for how they change during use, rather than just how they look when they are new. As the team looks ahead, they plan to use this same automated approach to explore even wider ranges of materials, aiming to find solutions that are not only effective but also sustainable and scalable for the global energy transition.