A machine learning-based framework for mapping hydrogen at the atomic scale

Qingkun Zhao,Qi Zhu,Zhenghao Zhang,Binglun Yin,Huajian Gao,Haofei Zhou

DOI: https://doi.org/10.1073/pnas.2410968121

2024-09-24

Abstract:Hydrogen, the lightest and most abundant element in the universe, plays essential roles in a variety of clean energy technologies and industrial processes. For over a century, it has been known that hydrogen can significantly degrade the mechanical properties of materials, leading to issues like hydrogen embrittlement. A major challenge that has significantly limited scientific advances in this field is that light atoms like hydrogen are difficult to image, even with state-of-the-art microscopic techniques. To address this challenge, here, we introduce Atom-H, a versatile and generalizable machine learning-based framework for imaging hydrogen atoms at the atomic scale. Using a high-resolution electron microscope image as input, Atom-H accurately captures the distribution of hydrogen atoms and local stresses at lattice defects, including dislocations, grain boundaries, cracks, and phase boundaries. This provides atomic-scale insights into hydrogen-governed mechanical behaviors in metallic materials, including pure metals like Ni, Fe, Ti and alloys like FeCr. The proposed framework has an immediate impact on current research into hydrogen embrittlement and is expected to have far-reaching implications for mapping "invisible" atoms in other scientific disciplines.