Operando spectral imaging of the lithium ion battery’s solid-electrolyte interphase

The lithium-ion battery is currently the preferred power source for applications ranging from smart phones to electric vehicles. Imaging the chemical reactions governing its function as they happen, with nanoscale spatial resolution and chemical specificity, is a long-standing open problem. Here, we demonstrate operando spectrum imaging of a Li-ion battery anode over multiple charge-discharge cycles using electron energy-loss spectroscopy (EELS) in a scanning transmission electron microscope (STEM). Using ultrathin Li-ion cells, we acquire reference EELS spectra for the various constituents of the solid-electrolyte interphase (SEI) layer and then apply these “chemical fingerprints” to high-resolution, real-space mapping of the corresponding physical structures. We observe the growth of Li and LiH dendrites in the SEI and fingerprint the SEI itself. High spatial- and spectral-resolution operando imaging of the air-sensitive liquid chemistries of the Li-ion cell opens a direct route to understanding the complex, dynamic mechanisms that affect battery safety, capacity, and lifetime.

: Imaging parameters. To ensure that the 'low-loss' spectra capture the entire zero-loss peak, we acquire an energy range of roughly −10 to 90 eV in those datasets. We write 'low-loss' because these spectra also capture the Li core-loss signal near 55 eV, which is at a particularly low energy due to lithium's small atomic number Z = 3 [42]. The 'core-loss' spectrum image that is acquired after the Fig. 4 low-loss spectrum image spans 11-2059 eV, which covers the carbon, nitrogen, and oxygen K-edges. This spectrum image is discussed in Figs S9-S12.

SM.2 Movie captions
Movie S1: Spectrum images of first lithiation, delithiated graphite, and second lithiation. This movie combines the data from the three spectrum images that are featured in Fig. 2 and Fig. 3. ADF images (top row), energy filtered images (second row, 50 meV bandwidth), and the full spectra for several reference materials (bottom, see Fig. 2) are shown together. The particular energy slice shown in the energy filtered image is indicated by the dotted line on the bottom plot. The frame shown in the still figure above this caption highlights the LiH signal, which appears strongly in dendrites in the lower left of the cycle 1 field of view and in the lower right of the cycle 2 field of view. To best explore the data set, a movie player with a status bar that can be dragged to highlight various energy windows of interest is recommended for viewing.
Movie S2: Annular dark field imaging of first delithiation. ADF STEM images of a graphite flake (left) in the process of delithiating are synchronized with the cyclic voltammogram (CV) (right). Here the potential is ramped at a constant rate of 2 mV/s. The 1024 × 1024 pixel ADF STEM images that constitute the movie are acquired with manual timing at an approximate rate of one frame per 30 s, where the frame acquisition time is 20 s and the time required to save each image is approximately 10 s. The video plays the 15 ADF images at a rate of 1 frame per second. To see the relatively small effect of imaging vs. lithiation/delithiation on this sample, see also Fig. 1 and Movie S3.
Movie S3: Annular dark field imaging of second delithiation. ADF STEM images of a graphite flake (left) in the process of delithiating are synchronized with the cyclic voltammogram (CV) (right). Here the potential is ramped at a constant rate of 2 mV/s. The first seventeen 1024 × 1024 pixel ADF STEM images (21 images in total constitute the movie) are acquired with manual timing at an approximate rate of one frame per 30 s, where the frame acquisition time is 20 s and the time required to save each image is approximately 10 s. The delays between the last four ADF images are longer (up to several minutes). To see the relatively small effect of imaging vs. lithiation/delithiation on this sample, see also Fig. 1 and Movie S2.
In this movie, and perhaps in Movie S2, some motion of the electrolyte is evident. This motion has little effect on the measured current (see particularly the time t > 10 s in this movie), which highlights another advantage of capping the electrodes with ALD alumina. The working electrode is immersed where it is bare, except perhaps over some small regions of graphite shown (the Pt current collector for the working electrode is covered with ALD alumina). Thus, electrolyte motion is not producing sizable electrical currents. The measured electrical currents are due to chemistry (intercalation, dendrite growth, SEI formation, etc.), as desired.
Movie S4: Spectrum image of well-developed SEI. This movie summarizes the spectrum image used to generate Figs. 4B-E and Fig. S8. The ADF STEM image shown on top is acquired simultaneously with the spectrum image. As with Movie S1, the scanning dotted line indicates the 50 meV energy slice shown in the EELS image. (Again, a movie player with a status bar that can be dragged to highlight various energy windows of interest is recommended for viewing.) The ADF image shows some changes from the previously acquired survey image (Fig. 4A). For example, a dark region at the far right, most likely a gas bubble, has evolved. Other signs of beam damage include the single-pixel-size spots in the spectrum image that are brighter than their surroundings at low energy and darker at high energy. We attribute these spots to chance overlap between the exact beam position and a location damaged by the beam previously (recall that the probe size is much smaller than the pixel size -see Table S1) during survey imaging. These spots appear only in the SEI and appear more frequently with repeated imaging, highlighting the particular sensitivity of the SEI to beam damage.  Figure S1: Fluid cell assembly diagram. The final steps of the cell assembly take place in a custom-made glove bag under an argon atmosphere. Inside the bag, three sides of the cell are epoxied together using the micromanipulators in a probe station equipped with an optical microscope. The fourth and final side is epoxied under vacuum (see Fig. 1 and Methods). The small, easily assembled vacuum chamber sits under the microscope objective and consists of a custom vacuum chuck, an O-ring, and a glass slide as shown (not to scale). Mechanical access for applying the epoxy is provided by a pair of 2-mm cube NdFeB magnets that transmit forces applied to the magnet outside the chamber, through the glass, to the magnet inside the chamber. The magnets slide freely on the glass, allowing the operator to apply the final, sealing coat of epoxy under vacuum, and thus to create a fluid cell that will remain thin in the TEM vacuum.  Figure S4: EELS thickness maps, first sample. These six images are constructed from the spectrum images used to generate Figs. 2-3. Created by integrating over the entire energy spectrum, the images in the left column give a 'bright field' representation of the sample, illustrating its structural evolution during cycle 1 and 2, and its (static) structure between the cycles. In the right column, the relative thickness maps are created using the 'log-ratio' method [50]. The intensity scale bars describe the dimensionless thickness ratio t/λ, where t is the sample thickness and λ is the electron mean free path (mfp). For the imaging conditions used, the following materials have an estimated [50] mfp, given in parentheses, of: lithium (200 nm), lithium hydride (220 nm), graphite (160 nm), Li 2 CO 3 (160 nm), EC:DMC (180 nm), and silicon nitride (140 nm). These mfp numbers, especially those of the compounds, are to be considered rough, as they are derived assuming standard elemental densities. By themselves, the two 20-nm-thick Si 3 N 4 membrane windows separating the fluid cell interior from the TEM vacuum represent a t/λ of 0.29, within 10% of the minimum total thickness measured. Taking a mfp of 200 nm to be characteristic of the other materials, we see that the non-window material varies from near zero to 70-, 60-, and 90-nm-thick for the cycle 1, before cycle 2, and cycle 2 spectrum images respectively. Interestingly, the cell thickness varies both in space and in time. For these spectrum images a t/λ ∼ 0.5 is typical, which corresponds to 40 nm of Si 3 N 4 and 30-50 nm of LIB material.
2.74 1.76 0.77 total intensity relative thickness 500 nm Figure S5: EELS thickness map, second sample. These two images are constructed from the spectrum image used to generate Fig. 4 using the methods of Fig. S4. This sample is thicker than the first: t/λ varies from 0.8 to 2.7, of which 0.5 to 2.4 can be attributed to the contents of the fluid cell. Again taking a mfp of 200 nm to be characteristic of the LIB materials, we see that the non-window material varies from 100 to 500-nm thick. Unlike the ADF STEM image (Fig. 4A), which captures electrons scattered to angles 20-40 mrad and highlights the Li and LiH dendrites within the SEI, the total intensity image highlights the SEI and dendrites uniformly. The relative thickness image shows the dendrites to have larger t/λ relative to the rest of the SEI.  Fig. 2B and Fig. 2E, respectively. Colored boxes in the right ADF image indicate regions that, due to their location and morphological and spectral uniformity, can be considered to be representative of a given material. (B) Representative spectra for Li, LiH, the SEI, the Si 3 N 4 & electrolyte, and the graphite, as determined by averaging the acquired spectra across the regions framed by correspondingly colored rectangular boxes in part (A). No subtractions are performed. For instance, the 'graphite' spectrum also includes a contribution from the Si 3 N 4 membrane window, which spans the entire field of view. MLLS fitting is applied over the energy range 4-40 eV. Li K-edge signals (right plots, which are zoom views of the left plots) show evidence of Li in the lithium, the lithium hydride, and the SEI, but not in the other materials, as expected. 12  Figure S8: Energy segmented spectrum image of a well developed SEI. To give a still rendition of Movie S4, the source spectrum image is segmented into 8 different energy slices, each of which corresponds to the signal integrated over the indicated energy bandwidth. The bandwidths vary from slice to slice. The frames with the red, green, and blue colored outlines show energy windows that indicate roughly the spatial distribution of the Li, LiH, and the SEI, respectively. This crude segmented display corroborates the more sophisticated MLLS approach presented in Fig. 4 of the main text.

MLLS map
Li K edge signal first spectum image Li K edge signal second spectrum image A B C D SEI LiH Li Figure S9: Li spectrum imaging, map summary. The ADF image (A), a reproduction of Fig. 4A, shows the sample before any spectrum imaging. The orange box indicates the field of view of the second, 'core-loss' spectrum image of the same sample (Table S1). The MLLS map (B) shows the spatial distribution of the lithium, the lithium hydride, and the SEI, as determined by fits to the low-loss portion of the first spectrum image using reference spectra acquired from the indicated regions. At a given pixel, integrating the individual background-subtracted core-loss intensities provides an indication of the quantity of corresponding element at that location. Applying this algorithm, which is described in detail in the caption of Fig. S11, to the Li K-edge signal produces (C) and (D) from the first (102.4 eV bandwidth) and second (2048 eV bandwidth) spectrum images, respectively. [The map (D) is a reproduction of Fig. S11C.] Relative to the Li K-edge signal in the lithium hydride dendrite, the Li K-edge signal in the lithium dendrite goes from being more intense in (C) to less intense in (D). Thus the lithium appears to be less stable than the lithium hydride under these (room temperature liquid cell) imaging conditions. The scale bars are 500 nm. Figure S10: 2048 eV bandwidth spectrum image, spectra. Average spectra collected from the correspondingly colored green and blue rectangles in the 11-2059 eV bandwidth spectrum image (Fig. S9). The insets show the Li K-edge (lower left), and the background-subtracted C, N, and O K-edges from the SEI (upper right), respectively, all from the same spectrum. K-edge resonances at 284 and 532 eV reveal the presence of carbon and oxygen, respectively, in the SEI. The nitrogen core-loss signal at 402 eV we attribute to the silicon nitride membrane windows. With this larger bandwidth the Li K-edge is still visible in the lithium hydride, but not in the SEI. The observed carbon and oxygen K-edge structure in the SEI is consistent with lithium carbonate and lithium semi-carbonates (i.e. alkyl carbonates) [8,10,30,34], but the signal-to-noise ratio and the confounding effects of the electrolyte background together prevent the stoichiometry or chemical bonding states from being quantified precisely (Fig. S11).  Table S1). (B) Annular dark field image acquired simultaneously. (C) Lithium signal after the Li K-edge at 55 eV, (D) carbon signal after the C K-edge at 284 eV, (E) nitrogen signal after the N K-edge at 402 eV, and (F) oxygen signal after the O K-edge at 532 eV. To extract the Li core-loss signal (C) we subtract an exponential background determined in a 20 eV bandwidth immediately below the edge, and integrate the remaining signal in a 11 eV bandwidth after the edge. To extract the C, N, and O core-loss signals (D,E,F), we subtract a power-law background determined in a 70 eV bandwidth 5-10 eV before the corresponding edge, and integrate the remaining signal in a 105 eV bandwidth after the edge. The lithium hydride dendrites show a significantly stronger Li signal than the lithium metal and some indication of an oxide shell [30]. The graphite is clearly visible at the top of all four (unnormalized) K-edge maps. Its presence in the nitrogen map, which might otherwise be expected to show a flat field, indicates that the K-edge intensities cannot be simply interpreted as proportional to the number of atoms of the corresponding element present.

Spectrum Image
Spectrum Image Spectrum Image A B C 1 μm Figure S12: Dose effects in survey images. Survey images acquired before the (A) first, (B) second, and (C) third spectrum images of the Fig. 4 sample. Acquiring the first spectrum image damages the pristine SEI (A), as evidenced by the decreased contrast difference between the Li and LiH dendrites and their surroundings (B). The second spectrum image, with its 10× longer dwell times (Table S1), imprints a visible grid of damage (C) in the SEI, but not elsewhere. Because the SEI is so obviously compromised after the second spectrum image, no data from the third spectrum image are presented in this paper. Improving the signal-to-noise ratio for spectroscopy of the SEI represents a major challenge, as the carbon-containing components of the SEI are more beam sensitive than the other cell components (i.e. the Li, the LiH, the graphite electrode, and the electrolyte) [31,33]. This sensitivity is evident in the first spectrum image (Movie S4), which shows single pixel spots that are noticeably brighter than their surroundings below 40 eV. We attribute these spots to chance overlap of the beam with a region previously damaged during survey imaging. They appear only in the SEI, and become more common with repeated imaging. Spectrum imaging, with its characteristically longer dwell times, makes the damage obvious even to ADF imaging by imprinting an array in the SEI that reflects the mismatch between the probe size and the pixel size (Table. S1). We conclude that, in its native, room-temperature liquid electrolyte, even comparatively lowdose survey imaging damages the SEI, and that the ability of any spectrum to faithfully reflect the SEI's constituent compounds in the unimaged, pristine state must be critically considered in light of the total dose delivered.   Figure S15: (Fig. 3, vector format) Chemical evolution at the solid-liquid interface during lithiation. ADF STEM images of the graphite flake with MLLS color overlays (see Fig. 2) (A) during the first lithiation, (B) before cycle 2, and (C) during lithiation 2 (left) and the corresponding, energy-filtered maps of the EELS intensity integrated over 50-80 eV, after background subtraction (right). We emphasize that, once the reference spectra are chosen, the MLLS mapping is entirely automated over the whole field of view. (D, E) Background-subtracted spectra summed over the regions indicated by red boxes in (A-C). Each box is 5 × 5 pixels, which corresponds to 107 nm×107 nm. The scale bar is 500 nm in (A-C). Corresponding grayscale MLLS images showing the individual components of (B, C), respectively (Fig. S7). The scale bars are 500 nm. (F) Background-subtracted Li core-loss signals found by calculating intensityweighted averages over the maps (D). 22