Supplementary information files for "Seeing the middle: reconstructing 3D internal electrode microstructures from low-resolution surfaces with generative diffusion artificial intelligence"

Supplementary files for article "Seeing the middle: reconstructing 3D internal electrode microstructures from low‐resolution surfaces with generative diffusion artificial intelligence"<br><br>Characterizing the 3D complex energy materials interface is critical to understand the correlative relationship between performance, degradation, and structures. Unfortunately, the resolution of microscopy and image acquisition speed are limited by the nature of the hardware, causing high-throughput characterization of energy materials to be prohibitive. Herein, REMind, a generative diffusion artificial intelligence model for fast and accurate reconstruction of electrode microstructures via focused ion beam-scanning electron microscopy, is presented. REMind can generate high-resolution internal microstructures between two low-resolution surfaces after training on sufficient high-resolution microstructures, enabling larger milling thickness between slices while keeping high-fidelity imaging. REMind is first demonstrated for reconstructing solid oxide fuel cell (SOFC) anode microstructures. REMind resolves relevant multi-scale structures with low pixel-wise reconstruction error (&lt;10%) and quantifies the generated uncertainty by calculating the generated entropy. Additionally, a multi-scale multi-physics SOFC model is employed to further quantify the reconstructed error regarding the electrochemical performance, i.e., operating current density versus overpotential. REMind shows good transferability, as proven by its ability to reconstruct other energy materials, including catalyst layers of proton exchange membrane fuel cells and solid-state battery composite electrodes, demonstrating the potential for REMind to be used as a general-purpose platform for broad development of energy technology.<br>©The Author(s), CC BY 4.0

论文《"透视中间态：基于生成式扩散人工智能从低分辨率表面重建三维电极内部微观结构"》补充材料

表征三维复杂能源材料界面，对于揭示性能、衰减与结构之间的关联关系至关重要。然而，显微镜分辨率与图像采集速度受硬件固有特性限制，使得能源材料的高通量表征难以开展。本文提出REMind模型：一种基于聚焦离子束-扫描电子显微镜的生成式扩散人工智能（generative diffusion artificial intelligence）模型，可实现电极微观结构的快速精准重建。REMind在充足的高分辨率微观结构数据集上完成训练后，可基于两个低分辨率表面生成高分辨率内部微观结构，能够在保持成像高保真度的前提下，增大切片间的铣削厚度。

该模型首先被应用于固体氧化物燃料电池（solid oxide fuel cell, SOFC）阳极微观结构的重建。REMind可实现相关多尺度结构的重建，且逐像素重建误差低于10%，并通过计算生成熵量化了生成结果的不确定性。此外，本文采用多尺度多物理场SOFC模型，进一步从电化学性能——即工作电流密度与过电位的对应关系——角度量化了重建误差。

REMind展现出优异的迁移性能：其可成功重建质子交换膜燃料电池催化层、固态电池复合电极等多种能源材料的微观结构，证明REMind有望成为支撑能源技术广泛发展的通用平台。

©作者，CC BY 4.0许可