At Tohoku University's Advanced Institute for Materials Research, Hao Li and an international team have launched DigMethpy, an artificial intelligence platform designed to solve one of clean energy's most stubborn puzzles: how to discover the right molten catalysts for methane pyrolysis at scale.
The challenge is urgent. Hydrogen is essential to our clean-energy future, yet most current production methods release carbon dioxide as waste. Methane pyrolysis offers a fundamentally different path—it splits methane into hydrogen and solid carbon, eliminating direct carbon emissions in the process. The catch: finding catalysts efficient enough to make the reaction work has required scientists to wade through an enormous and poorly understood chemical design space, relying on the slow, expensive tradition of trial-and-error experimentation.
DigMethpy changes that equation by weaving together scientific literature, experimental data, computational simulations, machine-learning models, and large language models into a single discovery framework. The platform already contains more than 40,000 carefully curated data points drawn from more than 500 scientific publications and computational records—covering molten metals, alloys, salts, and mixed catalyst systems. Rather than stopping there, DigMethpy operates as a closed-loop system: it continuously gathers information, predicts the most promising catalyst candidates, and refines its recommendations through validation feedback, getting smarter with each iteration.
Using this approach, researchers identified key chemical properties that determine catalyst performance—atomic charge-related descriptors, diffusion behavior, and hydrogen adsorption characteristics among them. These insights guided the design of highly active multicomponent molten alloy catalysts specifically tailored for methane pyrolysis. The results represent not just incremental progress but a shift in how materials research itself happens, turning the traditional laboratory bottleneck into a data-driven advantage.
Li, who serves as founding editor of the journal AI Agent where the research was published, sees this as the beginning of something larger. "By connecting experimental knowledge, computational modeling, machine learning, and large language models in a unified workflow, we can accelerate the development of catalysts needed for cleaner hydrogen production and other sustainable energy technologies," he explained. The achievement demonstrates how artificial intelligence can be meaningfully integrated into materials science to support faster, more reliable decision-making across the entire discovery pipeline.
The practical implications ripple outward. Scientists can now make better use of the explosion of available research data while dramatically cutting the time and expense of finding new catalytic materials. For hydrogen production, which will anchor the decarbonized energy systems of the coming decades, that acceleration matters enormously. Every catalyst discovered faster is another step toward making clean hydrogen economically competitive with fossil fuel alternatives.
The team plans to expand the platform further—growing the DigMethpy database, sharpening its predictive accuracy, and developing autonomous multi-agent systems that could eventually design catalysts with minimal human intervention. As the boundaries between computational discovery and experimental validation continue to blur, platforms like DigMethpy point toward a future where the slowest part of materials innovation isn't the science itself, but how quickly we can ask the right questions of the data we've already gathered.