Lithium Metal Battery Quality Control via Transformer-CNN Segmentation.

Area of Science:

• Materials Science
• Electrochemistry
• Artificial Intelligence

Background:

• Lithium metal batteries (LMBs) offer high energy density but are hindered by lithium dendrite formation.
• X-ray computed tomography (XCT) is crucial for non-destructive observation of dendrite morphology.
• Accurate 3D reconstruction of dendrites from XCT data requires robust image segmentation.

Purpose of the Study:

• To introduce TransforCNN, a novel transformer-based neural network for semantic segmentation of lithium dendrites in XCT data.
• To quantitatively compare the performance of TransforCNN against existing segmentation algorithms (U-Net, Y-Net, E-Net).

Main Methods:

• Development and application of a transformer-based neural network (TransforCNN) for semantic segmentation.
• Comparative analysis of TransforCNN with U-Net, Y-Net, and E-Net using XCT data from lithium metal batteries.
• Evaluation using over-segmentation metrics like mean intersection over union (mIoU) and mean Dice similarity coefficient (mDSC).

Main Results:

• TransforCNN demonstrated superior performance in segmenting lithium dendrites compared to U-Net, Y-Net, and E-Net.
• Quantitative evaluation showed significant advantages for TransforCNN in mIoU and mDSC metrics.
• Qualitative visualizations confirmed the effectiveness of TransforCNN in capturing intricate dendrite structures.

Conclusions:

• TransforCNN offers a powerful and accurate solution for segmenting lithium dendrites in XCT images.
• This advancement facilitates better quantitative analysis and understanding of dendrite formation in LMBs.
• The proposed method has the potential to accelerate the development and commercialization of safer and more efficient LMBs.