A Sandia team, as part of the Department of Energy HydroGEN consortium and in collaboration with the National Laboratory of the Rockies and Lawrence Livermore National Laboratory, recently developed a new artificial intelligence and machine learning (AI/ML) model to identify new materials that can be used in energy applications, like producing hydrogen from water using heat, converting CO2, or identifying radiation-hardened materials for electronics. The new technique is based on graph neural networks and is used to accelerate predictions of how much energy is needed to create vacancies, or empty spots, in different crystal structures. Compared to the computationally expensive first-principles methods on which the new model is trained, this model significantly speeds up the process of identifying materials with desired thermodynamics for vacancy formation, allowing researchers to quickly analyze thousands of crystal structures that may require millions of unique vacancy defect formation energy predictions.
“This new method of screening defect properties in solid state materials will help us significantly with materials discovery. So far, we’ve only applied it to thermochemical hydrogen (TCH) production materials, but by training the ML model on more diverse chemistries and structures, we could eventually apply it to more spaces,” explained Sandia’s Matt Witman.
By screening materials from existing databases, the new method can identify promising candidates much faster than before and link the predicted vacancy defect thermodynamics to real-world conditions, helping researchers identity the best materials for various applications. The model is extensible, with the team recently demonstrating its ability to predict vacancy migration energies (how fast they move in a given material); and the model is continuously improving as it is gradually trained on more collected data, making it even more effective for discovering new materials.
Most recently, the team expanded their solid-state material defect work by experimentally validating the model’s predictions of metal oxides that might perform well in TCH based on their oxygen vacancy properties. The team selected twelve material candidates and found that about 80 percent of them were effective in producing hydrogen during experiments.
“Some of these materials are comparable to or better than the current state-of-the-art material under certain reactor operating conditions. We found highly promising candidates to continue detailed investigation and identify modification strategies or similar materials to continue to improve performance,” shared Matt.
By using more precise calculations for training the AI/ML model, the team is confident that the accuracy in identifying experimentally active water-splitters will continue to improve. Key findings from this research were published in a recent article , most notably new materials the team discovered to produce more hydrogen in more commercially relevant conditions than state-of-the-art materials.
The HydroGEN consortium is a group of U.S. Department of Energy national laboratories — including National Laboratory of the Rockies, Lawrence Berkley National Laboratory, Sandia National Laboratories, Idaho National Laboratory, and Lawrence Livermore National Laboratory — focused on advancing water-splitting technologies for hydrogen production.
December 15, 2025