Deep-Learning Aided Atomic-Scale Phase Segmentation toward Diagnosing Complex Oxide Cathodes for Lithium-Ion Batteries

Phase transformation-a universal phenomenon in materials-plays a key role in determining their properties. Resolving complex phase domains in materials is critical to fostering a new fundamental understanding that facilitates new material development. So far, although conventional classification strategies such as order-parameter methods have been developed to distinguish remarkably disparate phases, highly accurate and efficient phase segmentation for material systems composed of multiphases remains unavailable. Here, by coupling hard-attention-enhanced U-Net network and geometry simulation with atomic-resolution transmission electron microscopy, we successfully developed a deep-learning tool enabling automated atom-by-atom phase segmentation of intertwined phase domains in technologically important cathode materials for lithium-ion batteries. The new strategy outperforms traditional methods and quantitatively elucidates the correlation between the multiple phases formed during battery operation. Our work demonstrates how deep learning can be employed to foster an in-depth understanding of phase transformation-related key issues in complex materials.

Automated summary

Phase transformation is a universal phenomenon in materials and plays a crucial role in determining their properties. However, accurately segmenting different phases in complex material systems remains a challenge. This study successfully developed a deep-learning tool that can automatically segment intertwined phase domains in cathode materials for lithium-ion batteries, outperforming traditional methods and providing quantitative insights into the correlation between multiple phases formed during battery operation.