Abstract
AbstractAccurate 3D representations of lithium-ion battery electrodes, in which the active particles, binder and pore phases are distinguished and labeled, can assist in understanding and ultimately improving battery performance. Here, we demonstrate a methodology for using deep-learning tools to achieve reliable segmentations of volumetric images of electrodes on which standard segmentation approaches fail due to insufficient contrast. We implement the 3D U-Net architecture for segmentation, and, to overcome the limitations of training data obtained experimentally through imaging, we show how synthetic learning data, consisting of realistic artificial electrode structures and their tomographic reconstructions, can be generated and used to enhance network performance. We apply our method to segment x-ray tomographic microscopy images of graphite-silicon composite electrodes and show it is accurate across standard metrics. We then apply it to obtain a statistically meaningful analysis of the microstructural evolution of the carbon-black and binder domain during battery operation.
💻 Software Details
🏛️ Research Organizations (ROR)
Affiliated research institutions:
📋 Methods
Electrode preparation
The electrode slurry is prepared by mixing 75 wt.% graphite (SLP 30 Timcal) with 10 wt.% silicon (BASF, SiO x , x ≈ 1), 10 wt.% polyvinyldiene difluoride (PVDF) binder (Kynar Flex® HSV900), and 5 wt.% carbon black (Timcal Super C60, Imerys). To improve the imaging contrast, 50 vol.% of the carbon black in the carbon black-binder domain is replaced by copper nanoparticles (US Research Nanomaterials, Inc.) 8 , 22 . The slurry is coated on copper foil with a doctor blade (150 μm blade gap), and dried overnight at 80 °C with nitrogen flow. 13-mm disk-shaped electrodes are punched out and compressed with 1 t. Further processing takes place in an Argon-filled glovebox. Electrochemical cycling A half-cell containing electrode, a 250 μm thick glass fiber separator (Whatman® glass microfiber filter), 500 μL standard LP50 electrolyte (BASF, 1 M LiPF6 in ethylene carbonate: ethyl methyl carbonate = 1:1 by weight) and lithium metal as a counter electrode (Alfa Aesar, lithium foil, 99.9%) is assembled in an Argon-filled glovebox. The cells are cycled galvanostatically in a potential range of 10 mV–1.5 V at a rate of C/20 for 2, 5, and 8 cycles with VMP3 battery cycling system (Biologic). Each protocol ends with a 20 h period of constant voltage (3 V).
Sample preparation for imaging
Samples with a diameter of 1 mm are punched out of the electrode and glued to a custom-made invar sample holder. Laser milling reduces the sample diameter to below 70 μm. XTM measurements All XTM measurements are recorded at the 32-ID-C beamline of the Advanced Photon Source at the Argonne National Laboratory at a beam energy of 8.4 keV. 1210 projections per tomographic scan are acquired with an exposure time of 2 s resulting in a measurement time of ∼40 min. The 2448 × 2448 pixel detector leads to a voxel size of 27.5 nm. Reconstruction The ASTRA toolbox 46 in MATLAB is used to filter and reconstruct the acquired projections with the Filter Back Projection (FBP) algorithm. Paganin phase retrieval 68 was performed with a ratio of the refractive index parameters δ/β = 0.028 and δ/β = 0.28. Manual segmentation Manual Segmentation is performed using the Dragonfly software. More details can be found in Supplementary Note 5 .
Show full methods section
Electrode preparation
The electrode slurry is prepared by mixing 75 wt.% graphite (SLP 30 Timcal) with 10 wt.% silicon (BASF, SiO x , x ≈ 1), 10 wt.% polyvinyldiene difluoride (PVDF) binder (Kynar Flex® HSV900), and 5 wt.% carbon black (Timcal Super C60, Imerys). To improve the imaging contrast, 50 vol.% of the carbon black in the carbon black-binder domain is replaced by copper nanoparticles (US Research Nanomaterials, Inc.) 8 , 22 . The slurry is coated on copper foil with a doctor blade (150 μm blade gap), and dried overnight at 80 °C with nitrogen flow. 13-mm disk-shaped electrodes are punched out and compressed with 1 t. Further processing takes place in an Argon-filled glovebox. Electrochemical cycling A half-cell containing electrode, a 250 μm thick glass fiber separator (Whatman® glass microfiber filter), 500 μL standard LP50 electrolyte (BASF, 1 M LiPF6 in ethylene carbonate: ethyl methyl carbonate = 1:1 by weight) and lithium metal as a counter electrode (Alfa Aesar, lithium foil, 99.9%) is assembled in an Argon-filled glovebox. The cells are cycled galvanostatically in a potential range of 10 mV–1.5 V at a rate of C/20 for 2, 5, and 8 cycles with VMP3 battery cycling system (Biologic). Each protocol ends with a 20 h period of constant voltage (3 V).
Sample preparation for imaging
Samples with a diameter of 1 mm are punched out of the electrode and glued to a custom-made invar sample holder. Laser milling reduces the sample diameter to below 70 μm. XTM measurements All XTM measurements are recorded at the 32-ID-C beamline of the Advanced Photon Source at the Argonne National Laboratory at a beam energy of 8.4 keV. 1210 projections per tomographic scan are acquired with an exposure time of 2 s resulting in a measurement time of ∼40 min. The 2448 × 2448 pixel detector leads to a voxel size of 27.5 nm. Reconstruction The ASTRA toolbox 46 in MATLAB is used to filter and reconstruct the acquired projections with the Filter Back Projection (FBP) algorithm. Paganin phase retrieval 68 was performed with a ratio of the refractive index parameters δ/β = 0.028 and δ/β = 0.28. Manual segmentation Manual Segmentation is performed using the Dragonfly software. More details can be found in Supplementary Note 5 .
Generation of artificial learning data
Artificial learning data are generated using MATLAB, the style transfer algorithm CycleGAN and the ASTRA toolbox to simulate the tomography. Details are described in Supplementary Note 3 and Note 4 . Renderings Images and renderings are produced with Arivis, ImageJ, and Inkscape.
Structural analysis
The analysis of the 12 segmented datasets is performed in MATLAB using code from Legland et al. 69 . The diffusivity in the through-plane direction was calculated on volumes scaled by a factor of 0.5 (384 × 384 × 384 voxel) using the DiffuDict toolbox of the GeoDict2020 Software (Math2Market GmbH, Kaiserslauten, Germany) applying symmetric boundary conditions (Dirichlet). Both border planes (along the through-plane direction) are set to have a constant concentration which decreases edge effects. Symmetric boundary conditions are also applied in all tangential directions and the Neumann (zero flux) condition is used at the domain boundaries. The concentration drop across the through-plane direction was set to 1 and the (dimensionless) Laplace equation is solved in an iterative process by finite volume-based solver (EJ) and stops if the process becomes stationary (tolerance of 0.001).
Supplementary information Supplementary Information Peer Review File
📊 Figures
Fig. 1
Deep learning segmentation of battery electrodes.
The goal of this work is to demonstrate unsupervised, learning-based segmentation of complex volumetric datasets that cannot be easily segmented using standard techniques (e.g., thresholding). We work...
Fig. 2
Challenging segmentation.
( a ) A subsection of a tomogram of a graphite-silicon composite electrode imaged with XTM. Scale bar is 5u2009u03bcm. Neither ( b ) threshold- nor ( c ) random walker-based segmentation yield satisfy...
Fig. 3
Real learning data.
( a ) Ptychographic x-ray computed tomography (PXCT) and ( b ) x-ray tomographic microscopy (XTM) is performed on the graphite-silicon composite electrodes. Data from these two imaging approaches are ...
Fig. 4
Benefits of artificial learning data for segmentation.
A through-plane cross-section ( a ) and in-plane cross-section ( b ) of the raw data are shown in order to demonstrate the segmentation achieved using different training datasets. When trained only on...
Fig. 5
Generation of artificial learning data.
An initial basic structure is generated based on polygon shapes as shown in 2D ( a ) and 3D ( b ). Using a style transfer algorithm, a synthetic structure is achieved ( c , d ). Tomography simulations...
Fig. 6
Deep Learning Segmentation.
Illustration highlighting that for each sample, the deep learning approach is applied to 775 subvolumes ( a ), which, once segmented, are reassembled. The multiphase segmentation ( b ) for pristine sa...
Fig. 7
Analysis of the segmented electrodes.
Volume fractions ( a ) occupied by pore space (black), graphite particles (gray), carbon black-binder domain (CBD) (yellow), and silicon particles (blue) for the three pristine (pr.) and the three sam...
Fig. 8
Evolution of microstructure with cycling.
Reconstruction of a subvolume ( a ) of a pristine electrode showing how the carbon black-binder domain (CBD) (gold) clusters in pores, particularly around silicon particles (blue). Illustration showin...
Figure images are served from the NIH/NLM PubMed Central Open Access Subset or Europe PMC; copyright remains with the publishers and authors.
💬 Discussion
0 commentsNo comments yet. Be the first to start a discussion!
Leave a Comment