Hydrogen could help store renewable energy and power fuel cells, but storing it remains difficult because hydrogen gas has a low density under everyday conditions. Compressing it to high pressure or cooling it into a liquid requires substantial energy and specialized equipment. Solid materials that absorb or adsorb hydrogen offer an alternative, but finding materials that store enough hydrogen, release it under useful conditions, work quickly, and survive repeated use remains a complex challenge.
"Reliable AI-guided discovery in this field cannot come from faster predictions alone," said Seong-Hoon Jang, Associate Professor at Tohoku University. "It requires physical constraints, transparent data, and continuous feedback from real experiments built into the process from the start."
Artificial intelligence can search vast numbers of candidate materials, but a fast prediction is not necessarily a reliable one. Hydrogen-storage data are scattered across computational databases and scientific papers, and important details about sample preparation, measurement conditions, uncertainty, and unsuccessful experiments are often missing. Models trained on incomplete records can recommend materials that are physically unrealistic, difficult to synthesize, or unsuitable under practical operating conditions.
An international team led by Tohoku University has outlined a roadmap for making AI-guided hydrogen-storage research more trustworthy and useful. Drawing on discussions among 69 researchers spanning hydrogen-storage materials, AI, computational science, and self-driving laboratories, the team proposes a physics-aware ecosystem connecting four elements: reproducibility-aware data, models constrained by thermodynamics and kinetics, AI-driven inverse design, and experimental validation. In this framework, physical consistency, uncertainty, data provenance, and experimental feedback are built into every stage, rather than checked only after an AI has proposed a candidate.
The framework turns materials discovery into a learning cycle. AI proposes candidates, automated systems synthesize and measure them, and the results are returned to the database and model to inform the next decision. The authors also describe a longer-term concept they call a "digital twin," a virtual representation that stays synchronized with real experiments. This could help researchers detect model errors or material degradation and choose the next experiment where it will reduce uncertainty most effectively. The team highlights the Digital Hydrogen Platform (DigHyd), which organizes more than 30,000 thermodynamic entries from over 4,000 publications, as an example of the data foundation this approach requires.
This work is a perspective rather than the report of a newly discovered storage material. It identifies the system-level changes needed to move the field from trial and error toward research that is reproducible, adaptive, and able to learn continuously. Experimental validation and expert judgment remain essential, particularly because AI-generated candidates may not be synthesizable.
"The limiting factors have not been a lack of computing power," Jang added. "They have been fragmented data, incomplete experimental context, weak physical consistency, and poor feedback between prediction and experiment. Addressing these could make the search for safer, more compact, and more practical hydrogen storage faster and more reliable, supporting future clean-energy systems."
Next steps for the project include standardizing how hydrogen-storage data and experimental conditions are recorded, expanding information on kinetics, cycling, and degradation, and connecting physics-grounded models with AI design tools. Tohoku University is also integrating these tools with automated synthesis and measurement systems and is developing autonomous experimental workflows under the JST GteX program (JPMJGX23H1).
The perspective was published online in ACS Energy Letters on August 14, 2026.
- Publication Details:
Title: Building a Physics-Aware AI Ecosystem for Solid-State Hydrogen Storage Materials
Authors: Seong-Hoon Jang, Yiwen Yao, Chuanyu Liu, Linda Zhang, Di Zhang, Xue Jia, Hung Ba Tran, Eric Jianfeng Cheng, Ryuhei Sato, Yusuke Ohashi, Toyoto Sato, Yusuke Hashimoto, Mark D. Allendorf, Nongnuch Artrith, Marcello Baricco, Andreas Borgschulte, Darren P. Broom, Ang Cao, Benjamin Wei Jie Chen, Lixin Chen, Ping Chen, Eun Seon Cho, Stefano Deledda, Zhao Ding, Martin Dornheim, Michael Felderhoff, Yaroslav Filinchuk, George E. Froudakis, Mingxia Gao, Thomas Gennett, Zaiping Guo, Ikutaro Hamada, Jason Hattrick-Simpers, Bjørn C. Hauback, Michael Hirscher, Torben R. Jensen, Baohua Jia, Hyoung Seop Kim, Takahiro Kondo, Kentaro Kutsukake, Xiao-Yan Li, Tongliang Liu, Piao Ma, Jianfeng Mao, Rana Mohtadi, Hyunchul Oh, Mark Paskevicius, Chris J. Pickard, Astrid Pundt, Anibal J. Ramirez-Cuesta, Hiroyuki Saitoh, Kaihang Shi, Aloysius Soon, Chenghua Sun, Chris Wolverton, Hiroshi Yabu, Weijie Yang, Zhenpeng Yao, Xuebin Yu, Jianxin Zou, Shouyi Hu, Panpan Zhou, Xi Lin, Zhigang Hu, Zhenhao Zhou, Pengfei Ou, Jiayu Peng, Shin-ichi Orimo*, and Hao Li
Journal: ACS Energy Letters
