Probing lattice defects in crystalline battery cathode using hard X-ray nanoprobe with data-driven modeling

This repository contains the source codes of the following paper:

@unpublished{li2021nanodiffraction,
  title={Probing lattice defects in crystalline battery cathode using hard 
X-ray nanoprobe with data-driven modeling},
  author={Li, Jizhou and Hong, Yanshuai and Yan, Hanfei and Chu, Yong S. and Pianetta, Piero and Li, Hong and Ratner, Daniel and Huang, Xiaojing and Yu, Xiqian and Liu, Yijin},
  year={2021},
  note={doi: 10.1016/j.ensm.2021.12.019}
}

Background & Key Contributions

• Lattice defects, e.g., dislocations and grain boundaries, critically impact the properties of crystalline battery cathode materials.
• A long-standing challenge is to probe the meso-scale heterogeneity and evolution of lattice defects with sensitivity to atomic-scale details.
• We tackle this issue with a unique combination of X-ray nanoprobe diffractive imaging and advanced machine learning techniques.
• These results pave a direct way to the understanding of crystalline battery materials’ response under external stimuli with high fidelity, which provides valuable empirical guidance to defect-engineering strategies for improving the cathode materials against aggressive battery operation.

Imaging technique

The hard X-ray nanoprobe beamline 3-ID of National Synchrotron Light Source II is used to perform the scanning X-ray microscopy measurements. The experiment was performed with a 30x30 nm² X-ray beam. The scanning step size was 100 nm, thus the pixel size of the obtained images is 100x100 nm².

The sample was rotated over a 180^o range with an XRD detector recording the diffraction pattern at each rotation angle in order to locate the target Bragg peak. A pixel array detector was then oriented to measure the strongest (101) peak. The crystal was rocked over a 2^o angular range in the vicinity of the (101) Bragg peak, and a two-dimensional raster scan was conducted at each rocking angle. The local Bragg diffraction measurements were performed in sync with the raster scan, and the raster scans were repeated for a series of rocking angles with diffraction signals above the noise level.

Fig.1. Example diffraction patterns from the hard X-ray nanoprobe.

Imaging Data Analysis

• The raw diffraction pattern in our experiments at each pixel is a movie of about 0.5million pixels.
• The dimension for meaningful analysis is too high for conventional approaches.
• We thus have the following analysis pipeline:
  • Conversion to point clouds
  • Denoising
  • Feature extraction

Fig.2. Example of the converted point cloud of the nanodiffraction pattern.

Denoising

We use the LPA-ICI algorithm to remove the noise and strengthen the shape of point cloud data.

[pcd_de] = PCD_Filtering_LPAICI_mex(ptCloudB.Location,5，2);

Feature Extraction

The implementation is based on the FoldingNet approach. The autoencoder is trained by minimizing a Chamfer distance function between the original point-cloud data and reconstructed data. The encoder uses a graph-based architecture, and the decoder is based on the folding architecture, which essentially forms a universal 2D-to-3D mapping.

Fig.3. Illustration of the network architecture.

python main.py --exp_name diffractiontest --dataset_root /data --encoder foldnet --k 32 --dropout 0.5 --shape plane --dataset diffraction  --workers 50 

An example of applying the trained model to interprete the nanodiffraction data can be found here.