Introduction

The reactive nature of lithium (Li) represents also its strength and has lead about 50 years ago to the first functional lithium ion battery (LIB)1,2. The trend toward electrification, especially driven by electric vehicles, renewable energy storage and portable electronics, has highlighted the central importance of LIBs. Though, the ongoing electrification of society demands enhancements in energy densities, cycling stability and endurance for future Li-ion batteries. Here, material science starts to play a pivotal element in developing superior electrode materials, crucial for the next-generation Li-ion batteries. The integration of silicon (Si) with a high weight percentage into LIB anodes holds immense potential to revolutionize the energy storage landscape. The material not only presents a remarkable theoretical specific capacity (3579 mAh g−1 for Li15Si4), exceeding traditional graphite anodes by over tenfold, but it also stands out for its cost-effectiveness, natural abundance as well as its non-toxic behaviour3,4,5,6,7,8. Hence, the utilization of crystalline Si has been identified as a promising material, not just for anodes in Li-ion batteries9,10,11,12, but also highly relevant to emerging technologies like all-solid-state-batteries13,14,15,16,17.

With the implementation of Si in LIB anodes, challenges arise due to its recurrent volume expansion and shrinkage during the insertion and extraction of Li-ions. This lithiation mechanism is based on amorphization and conversion of silicon into a high-lithium-content LixSiy alloy5,18,19,20. The interaction between Li-ions, active material and electrolyte is giving rise to the formation of an interlayer known as the solid electrolyte interphase (SEI). The huge volume swing strongly impacts the SEI formation at the silicon-electrolyte interface. As the silicon expands, the SEI layer fractures, exposing pristine silicon surfaces to the electrolyte5,10,21,22,23,24. As a consequence of electrochemical cycling, repeated SEI10,22,23 growth leads to the formation of a silicon electrolyte composite (SEC)25 at the interface of the Si. This gradual process degrades the Si as well as depletes active lithium and electrolyte, resulting in a reduction of the cell’s capacity5,10,21,22,23,25,26,27. Importantly, the volume changes generate mechanical stresses that impact the silicon domain and its proximity. In particular, this may lead to particle pulverisation, the disruption of both Li-ion and electron conduction pathways, the isolation of silicon particles as well as to Li-trapping10,20,21,22,23,24,25,27,28. Consequently, this renders them inactive within the anode structure.

Despite considerable advancements in understanding the electrochemical aspects of Si-based anodes, the material science perspective on the electrochemically induced volume swing and the impacts of the associated mechanical loading on the silicon´s crystalline structure remains insufficiently explored. This represents a critical knowledge gap crucial for advancing the development of the next generation of anode materials. Extensive theoretical models and experimental studies on fracture mechanisms have demonstrated, that covalent crystals, such as silicon, can undergo a transition from the crystalline to amorphous phase when subjected to compression29,30,31,32, surface scratching33, or nanoindentation34. Amorphization can manifest within shear bands, arising from a high density of defects that destabilize the crystalline lattice30,31,35,36. A mechanically induced phase transition of the crystalline Si within the anode would have direct consequences on the lithiation and SEC formation process. Within this context, fundamental questions arise: Is the formation of shear bands, leading to the transformation from a crystalline to an amorphous phase in Si-based anodes, feasible? Is the stress exerted on the crystalline silicon core during electrochemical cycling sufficient? Can insights derived from smaller-scale investigations effectively elucidate the failure mechanisms in realistic, large-scale electrodes utilizing Si particles? Furthermore, could this triggered phase transition contribute to an accelerated capacity loss, or does it potentially have a positive effect on the overall lifetime of high-energy-density Li-ion batteries?

Herein, we investigate the degradation behaviour of silicon-based anodes in Li-ion batteries in full-cell configuration up to prolonged electrochemical cycling, unveiling the emergence of amorphous shear bands within the crystalline silicon structure and their impact on the lithiation and SEC formation process. A qualitative correlation between the SEC formation process and capacity fading is established by quantifying the degradation of the microstructure upon electrochemical cycling and identifying the experienced varying mechanical conditions on the silicon. Thereof, we conduct advanced imaging techniques on various length scales such as 4D scanning transmission electron microscopy (STEM), field emission scanning electron microscopy (FESEM), synchrotron X-ray nano-tomography as well as incorporate machine learning (ML)-based microstructure analysis. Our study elucidates how, cyclic electrochemical loading prompts non-hydrostatic strain and yields to shear bands within the crystalline silicon. The deformation induced amorphization eventually backfires to the recurring lithiation and intensifies a non-uniform SEC growth at the silicon interface affecting the capacity. The lattice deformation emerges after only three cycles within the silicon domains close to the interface. Further dislocation emission evolves upon prolonged cycling, progressing finally into shear bands under the generated strain, stretching across the entire silicon domain. Strain mapping, by 4D STEM, provides insights into the varying mechanical loading conditions experienced by the silicon domains due to electrochemical cycling. The ML-based model purveys statistical information about the accompanied degrading of the anode microstructure upon electrochemical cycling in context to the diminishment of the pore-, and silicon phase fractions as well as the growth of the SEC. Complementary molecular dynamics modelling, focusing on the mechanical effects of volumetric changes on the silicon particles’ core during electrochemical cycling, support the empirical observations from multiscale structural and chemical experimental characterization. We argue that the stress exerted on the crystalline silicon core is sufficient to form shear bands upon electrochemical cycling. Finally, we hypothesize that the formation of the amorphous regions within the crystalline silicon core not only influence the ongoing lithiation process but also impede complete particle pulverisation.

This comprehensive approach emphasizes the importance of integrating both electrochemical and mechanical mechanisms, providing valuable insights into the determinants affecting the cycling stability of silicon-rich anodes. It highlights the necessity of investigating all concurrent processes during battery operation across various length scales, essential for the successful advancement of batteries with enhanced cycling stability and minimized capacity loss.

Results

Structural integrity of the anode and capacity loss

In this study, we employ a full cell configuration featuring silicon-rich anodes, as depicted schematically in Fig. 1a. The micro-porous anode is composed of a blend containing micron-sized crystalline silicon and graphite particles embedded within a carbon-binder domain (CBD). To ensure an enhanced electron transport, carbon nanotubes (CNTs) are integrated into the CBD. The porous network facilitates both the transport of Li+ -ions and the mitigation of stress generated during electrochemical cycling35,37,38,39. These electrodes feature a mass loading of 3.8 mg cm−2 with a capacity of 3.6 mAh cm−2. The anode is separated from the positive electrode by an electron blocking membrane. Li-ion transport is enabled by a fluorinated 1 molar solution of lithium hexafluorophosphate (LiPF6) in a 1:2 (v/v) mixture of fluoroethylene carbonate (FEC) and ethyl methyl carbonate (EMC) with 2 wt% vinylene carbonate (VC). The cathode employs a lithium-nickel-manganese-cobalt-oxide (NMC) composition. Additional specifics are outlined in the Methods section and Supplementary Note 1. To assess the influence of electrochemical loading on the anode microstructure, the cells undergo electrochemical cycling for up to 300 cycles at a charging rate (C-rate) of C/2.

Fig. 1: Experimental setup of the utilised Si-anode, electrochemical performance and microstructure degradation.
figure 1

a Schematic of the utilised Li-ion cell setup featuring a silicon/graphite anode, separator, NMC-based cathode and a copper and aluminium current collector for the anode and cathode, respectively (not drawn to scale). b 3D renderings of the ML segmented X-ray nano-tomography volumes imaged at 25 nm voxel size illustrating different phase repartitions between pristine anode and anode after 300 cycles. On the left, the pristine anode composed of silicon (red) and graphite (green) particles is illustrated. On the right, the anode after 300 cycles depicts microstructural degradation of Si (red) and the micro-pores (transparent). Graphite is represented in green. Additionally, the emerging solid electrolyte composite (SEC) is shown in blue, absent in the pristine electrode. c Statistical information, extracted from the segmented tomographic data for the pristine anode and after 300 cycles, indicating the evolution of the phase proportions and in particular the degradation of the anode material. d Capacity evolution over 300 electrochemical cycles, retaining 59% capacity at cycle 300. Rate capability checks are performed at 24, 124, 224 and 324 cycles, see dashed lines.

Full size image

Synchrotron X-ray nano-tomography scans are carried out at the ID16B beamline40 of ESRF at voxel sizes of 25, 50 and 100 nm. Further details are provided in the Methods section. The objective of the nano-tomography measurements is to quantify the degradation of the active material at scales ranging from micrometres to sub-micrometres, covering up to 300 cycles. To evaluate the structural changes exhibiting in the anode during electrochemical cycling, an accurate segmentation of the different phases is required. To overcome the limitations of conventional image analysis segmentation, a deep learning (DL)-based semantic segmentation approach is employed, as further detailed in the Methods section and in Supplementary Note 2. Figure 1b presents rendered DL segmented volumes for both the pristine anode and the anode after 300 cycles, depicting substantial microstructural changes, including the reduction of the porosity and the silicon volume. Previous studies5,10,21,22 have emphasised the detrimental impact of repetitive lithiation and delithiation on silicon particles, evident in a significant reduction in the volume fraction of silicon. As shown in Fig. 1c, the volume of Si diminishes from 58.5% in the pristine state to 13.0% at 300 cycles. This reduction is accompanied by a decrease in porosity from 35.4% in the pristine state to 7.8% after 300 cycles. Yet, the diminution is coupled with the formation of an additional phase, likely associated with the reformation of the SEI during electrochemical cycling, leading to a micron-sized SEC25. The evaluated volume fraction of the graphite remains almost unchanged, supporting the argument that the electrochemical cycling minimally impacts the morphology or density of the graphite41. Supplementary Figure 3 provides additional grey-scale reconstructed image data with a voxelsize of 25 nm. Figure 1d depicts the mean discharge capacity. Upon completion of the formation phase with three cycles, the cells exhibit a capacity retention of 99.6 ± 0.3%. Following an extended cycling period of 300 cycles, the cells maintain about 59.3 ± 0.4% of their original capacity, see details in the method section and Supplementary Note 1. Based on the structure-property relationship, a qualitative correlation can be established between the modification of the microstructure upon electrochemical cycling and capacity fading. The increase in the quantified SEC volume fraction during electrochemical cycling, coupled with the reduction in porosity and silicon, see Fig. 1c, suggests a capacity loss as depicted in Fig. 1d. Note that a manifold of degradation mechanism can cause capacity fading, therefore a detailed diagnosis is essential by investigating the structural modifications and the experienced varying mechanical conditions on the silicon.

Multiscale chemical and structural evolution of the anode upon electrochemical cycling

Herein a combination of electron microscopy and energy-dispersive X-ray spectroscopy (EDS) is employed to investigate the microstructural and chemical evolution of the Si core and its interface with higher resolution and contrast, shedding comprehensive light on the relationship between silicon interface deformation and observed capacity loss. Figure 2a presents FESEM-EDS images of the anode after the third cycle, illustrating the infiltration effect of the fluorine-rich electrolyte into the anode’s pore network. A discernible layer rich in fluorine (F), oxygen (O), and carbon (C) envelops the silicon (Si) domains, indicating the initiation of microstructural changes, notably the formation of SEC. This observation is strikingly contrasted with the pristine anode, where only binder material surrounds and connects silicon and graphite domains (refer to Supplementary Figure 4).

Fig. 2: Si-interface degradation and Si-core anomalies upon electrochemical cycling utilizing microstructure imaging and EDS elemental composition mapping.
figure 2

a, b Cross-sectional FESEM-EDS images of the anode after 3 and 300 cycles, respectively. The sequence from left to right comprises the backscattered electron FESEM images and the EDS-element composition mappings illustrating silicon (red), carbon (green), oxygen (blue) and fluorine (yellow). c Magnified (BSE) FESEM image of a Si-domain after 300 cycles in the x-y cross sectional plane revealing an inhomogeneous lithiation at the Si-interface with unequal dendrite growth. The Si-core displays two domains with bright and dark grey values. Bright domains are associated with significant dendrite growth at the Si-interface, as indicated by white boxes, while Si-domain interfaces linked to a dark domain exhibit no dendrite formation (indicated with blue boxes). Domains with bright grey values within the core are exemplarily highlighted with white dashed lines. d Cross-sectional dark field (DF)-STEM image after 300 cycles indicates different silicon domains and their surroundings with SEC in the z-x cross sectional plane. Two region of interests (ROIs) are designated as ROI 1 and 3, respectively. The silicon domain in the upper right (partly within ROI 1) corresponds to the domain illustrated in c. An additional Si domain is observed in the lower part of the image (partly within ROI 3). e Magnified cross-sectional bright field (BF)-STEM image of ROI 1 and EDS elemental composition mappings within the highlighted ROI 2 (dashed white box). The interface between Si and SEC is indicated by an orange dotted line within a selected area. The EDS mappings for silicon (red) and fluorine (yellow), respectively, indicate no elemental differences between the bright and dark Si grey value domains. f EDS mapping of ROI 3 for silicon (red) while the EDS mapping for ROI 4 (dashed white box) is performed for carbon (C), oxygen (O) and fluorine (F) with green, blue and yellow, respectively. Notably, C, O and F are only detected at the Si-interface. The scalebars correspond to 2 µm for a-b, to 500 nm for c-d and f. For e the scale bar is 200 nm.

Full size image

After 300 cycles, as depicted in Fig. 2b, fewer silicon domains are identifiable, surrounded by an expanded SEC matrix. Larger domains exhibit markedly pitted surfaces, while many smaller silicon domains appear to have undergone complete decomposition. These microstructural alterations become easily evident by secondary electron SEM imaging under a slightly tilted angle within Supplementary Figure 5. This underscores the repetitive SEI growth and subsequent dissolution of silicon domains24,25,42. Supplementary Figure 6 quantifies these structural changes based on EDS mappings, illustrating the decline of pore and silicon fraction in favour of SEC growth over cycling, along with insights into the mass composition of elements detected within the SEC. Dendrite-like SEC formation consumes active material and hampers both electron and Li-ion transport, thus increasing impedance. Remarkably, despite capacity loss, the cores of Si domains remain intact after 300 cycles, devoid of fluorine, suggesting incomplete lithiation within the silicon core. Statistical phase distribution analysis, utilising a convolutional neural network with a U-NET architecture for backscattered electron image segmentation, demonstrates a reduction of porosity from 46.5 ± 1.7% in the pristine state to 16.4 ± 3.5% after 300 cycles (refer to Supplementary Note 3 and 4). Although a slight deviation from the tomography’s analyses is noted (see Fig. 1c), the overall trend remains consistent. This deviation can be attributed to a smaller ROI and reduced statistical data availability of the FESEM measurements, as well as the difference in scale and resolution covered by the tomography scans.

Figure 2c shows a magnified backscatter electron (BSE) FESEM image of a Si-domain after 300 cycles, revealing dendritic growth on both micrometre and nanometre scales at the silicon interface. FESEM investigations reveal SEC layer thicknesses ranging from several nanometres to a maximum of approximately one micrometre after 300 cycles. Notably, some regions exhibit almost smooth surfaces devoid of dendritic growth (see blue boxes), suggesting practically inert regions during lithiation, resulting in minimal SEC formation, similar as previously shown by Haufe et al. 10. The observed non-uniform SEC formation at the Si-interface might impact the capacity fading. Grey value changes within the Si core are witnessed and reveal distinctive areas of brighter grey embedded in darker grey domains within the FESEM image. These bright lines traverse the particle, converging and widening near the surface, particularly where dendrites are observed (white boxed interfaces). Note, these grey value differences are not observed in the pristine sample (see Supplementary Figure 9).

Scanning transmission electron microscopy is employed to further investigate the Si-domain of the anode after 300 cycles, see also Supplementary Note 5. Figure 2d displays the cross-sectional dark field (DF)-STEM image showing two silicon domains and their surroundings. The silicon domain located in the upper right, partially within ROI 1, corresponds to the domain illustrated in c, however in z-x direction. The inward SEC growth toward the Si core, inducing gradual decomposition of the Si domain, is detected upon closer investigation of ROI 1 within Fig. 2e. The SEC manifests as dendrite-like nano-porous structures enriched with fluorine, oxygen and carbon, as supplementary shown in Supplementary Figure 10. STEM examinations on the dendrites from Supplementary Figure 10 reveal a mean dendrite pitch size of around 550 nm, consistent with the observed average SEC thickness. The porous dendrites exhibit local growth in plate-like structures, indicating incremental growth in line with the underlying SEI reformation process. The distance between the plate-like structures measures approximately 23 nm consistent with literature43,44, indicating a SEI thickness of approximately 20 nm. Notably, the observed inward growth of the SEI acts as a barrier to Li+ diffusion, resulting in an uneven Li-ion flux across the SEI45, which could contribute to the inhomogeneous formation of the SEC. Importantly, SEC growth is exclusively pronounced at the silicon interface, leading to the chemical caused amorphization and decomposition of the initially crystalline Si at the particle’s interface.

In the STEM images, grey value differences in the silicon are even more apparent than in the FESEM investigations. In the bright field (BF)-STEM images of Fig. 2e and Supplementary Figure 10, a dark grey silicon domain surrounding a brighter grey becomes visible. Note that these images are BF images, thus having inversed contrast. In Fig. 2e dendrites are predominantly observed in the direct vicinity of the dark grey regions of the Si core, rather than within the light grey areas. These observations suggest a potential bypassing of dendritic growth around the light grey domains. Detailed EDS element mapping, as illustrated in Fig. 2f, indicates that the core of the micron-sized particle is exclusively composed of silicon and electrolyte-associated elements (F, C, O) which are primarily detected at the silicon domain interface. This suggests that the variations in grey values within the silicon core are unlikely to be attributed to particle cracking or the penetration of other elements.

Phase Transformation within individual Si-domains upon electrochemical cycling

In Fig. 3, we delve deeper into the individual silicon domains and their microstructural changes upon electrochemical cycling. To this end, cross-sectional bright-field scanning transmission electron microscopy (BF-STEM) is employed. Figure 3 illustrates image data for the pristine anode, and after three and 300 cycles. In Fig. 3a, the BF-STEM image of the pristine anode reveals various well separated silicon domains embedded in the CBD. The Si interface is dendrite-free and the pristine silicon domains exhibit a uniform grayscale without discernible areas of varying grey values. For information on crystallographic orientation and potential misorientations, transmission Kikuchi diffraction (TKD) measurements are conducted. The associated inverse pole figure (TKD-IPF) representation provides insight into the crystallographic orientation of the individual silicon domains. The TKD-IPF reveals multiple distinct single-grain silicon domains with various orientations highlighted by different colours. Non-indexed areas in black within the IPF correspond to regions associated with pores and non-crystalline domains like the CBD. Further, the kernel average misorientation (TKD-KAM) shows for the pristine anode negligible misorientations for the silicon. The atomic structure is investigated employing high-resolution transmission electron microscopy (HR-TEM). The HR-TEM image and Fourier transform analysis illustrates an undisturbed atomic lattice with a crystalline structure (c-Si) for the pristine silicon.

Fig. 3: Microscopy analysis of the silicon structure after electrochemical cycling.
figure 3

Cross-sectional bright field scanning transmission electron microscopy (BF-STEM), transmission Kikuchi diffraction (TKD) analysis and high-resolution (HR-)TEM characterization of the silicon structure for a the pristine anode, the anode after b the third, and c the 300th cycle. Note, that for the third and 300th cycle, in contrast to the pristine anode, a single Si domain is illustrated. The sequence from left to right includes the STEM image, the inverse pole figure (TKD-IPF) presentation, the kernel average misorientation (TKD-KAM) mapping, and high-resolution HR-TEM image of the selected ROIs. The corresponding diffractograms shown as inset in the HR-TEM image assess the crystallinity of the Si domains. For the pristine electrode the Si is purely crystalline. After 3 and 300 cycles the formation of crystalline and amorphous domains is shown, marked with dashed lines. Dashed boxes for 3 cycles indicate non-registered features in the TKD-IPF but showing grey value changes in the BF-STEM and deviations in TKD-KAM. The scalebars for STEM, IPF and KAM images correspond to 200 nm.

Full size image

After three electrochemical cycles, the BF-STEM image in Fig. 3b shows a single silicon domain next to a pore (lower right) embedded in the CBD and indicates initial decomposition of the silicon microstructure. Significantly, darker grey value pixels embedded in the brighter matrix are visible near the interface of the single silicon domain and towards the core, as lines and areas. The single silicon domain shows an indexed region with identical crystal orientation (yellow) in the TKD-IPF, interrupted by darker pixels mainly at the interface. HR-TEM is performed for a region of interest approximately 200 nm away from the Si interface incorporating such darker grey value pixels. Here, the HR-TEM image and Fourier transform analysis reveal the presence of Si with a crystalline lattice and the formation of nanosized amorphous regions, indicated as c-Si and a-Si, respectively. The a-Si is associated with the darker grey value pixels in BF-STEM image. In general, both the inverse pole figure (TKD-IPF) and kernel average misorientation images (TKD-KAM) vividly confirm the observed disruptions in the crystalline structure, relegating to the degradation of the Si upon electrochemical cycling. Yet, not all grey value inhomogeneities depicted in the BF-STEM data or misorientations detected in the TKD-KAM  match with the TKD-IPF. In Fig. 3b, exemplarily such non-registered areas within the TKD-IPF data are highlighted as dashed boxes. These regions, devoid of variations in the TKD-IPF but displaying alterations in the misorientation mapping or BF-STEM, may signify initial stages of amorphization.

After 300 electrochemical cycles, a substantial transformation of the single silicon domain becomes apparent in Fig. 3c. The presented image provides an enlarged view of the silicon domain featured in Fig. 2d. The BF-STEM image indicates more extended disruptions within the silicon as shown in Fig. 3b. The corresponding TKD-IPF reveals indexed regions with identical crystal orientations (red) alongside with non-indexed band-like structures (black) traversing the silicon domain correlating well with the TKD-KAM representation. The non-indexed band-like structures (black) align precisely with the bands (dark grey) observed in the BF-STEM data. HR-TEM imaging and Fourier transform analysis in Fig. 3c for a representative region of interest, indicated as ROI in the BF-STEM, TKD-IPF as well as the TKD-KAM and comprising indexed as well as non-indexed regions reveal nanosized amorphous bands embedded in the crystalline silicon. Notably, according to the HR-TEM image the transitions between amorphous (a)-Si and crystalline (c)-Si are not sharply delineated but rather form a thin layer depicting a distinct grey value. The observed amorphization shown in Fig. 3 suggests a different origin compared to the amorphization witnessed at the silicon interface, which is triggered by lithiation of Si and the successive dendrite formation due to the continuous SEI reformation (see Supplementary Figure 10 and 11) upon electrochemical cycling.

The role of Li-insertion and extraction on dislocation and shear band formation

The crystalline-to-amorphization transition in the Si domains after the 300th delithiation is subjected to further analysis using detailed DF- and BF-STEM imaging, along with selected area electron diffraction (SAD), as depicted in Fig. 4a–c. Examination of Fig. 4a, b reveals complex deformation features within the silicon core, with a significant accumulation of dislocations. These discernible sharp bands, identified as a-Si, attain a maximum width of approximately 170 nm. In Fig. 4b, the BF-STEM image of the selected ROI from Fig. 4a, is captured under higher magnification. It provides a closer examination of the dislocations and associated low angle grain boundaries (LAGB). This observation suggests that the incipient plasticity is rendered by the formation of dislocations, likely triggered by the experienced volume swing. The data indicates that stress induced by Li-ion insertion and extraction leads not only to amorphization at the interface due to SEC formation and the conversion of c-Si into a-Si after delithiation of the high-lithium-content LixSiy alloy5,18,19,20, see Fig. 2e, but also to a localized amorphization within the silicon core domain, see Fig. 4a, b. The SAD pattern of the bright lines, labelled as I in Fig. 4b, corroborates the region’s amorphous nature, see Fig. 4c. Despite these deformations, the remaining sample domains exhibit only minor local mis-tilts, maintaining their single-crystalline nature, as evidenced by the consistent [100] zone axis across nano-diffraction patterns II, III, and IV in Fig. 4c. These minor local deviations are discernible from variations in the intensities of the reflections.

Fig. 4: TEM analysis to illustrate the deformation induced dislocation and associated generated strain upon cycling.
figure 4

a Dark field (DF-)STEM image of a silicon domain after 300 cycles showing, amorphous bands within the crystalline structure as well as dislocation accumulation. ROI 1 is indicated. b Magnified BF-STEM image (ROI 1) highlighting deformation bands and low-angle grain boundaries (LAGBs). The LAGBs are illustrated as yellow dashed lines. c Nano-diffraction patterns (SAD) corresponding to the locations marked in b demonstrate the presence of both, crystalline regions and amorphous bands. d DF-STEM image after 3 cycles. Different domains like pores, SEC and c-Si are marked with black, green and red, respectively. In addition, dislocation accumulation in particular at the Si-interface is identified. e Magnified area (ROI 2 in Supplementary Figure 13) identifies screw (blue)- and edge (red) dislocations along the 111 plane within the crystalline c-Si. f Schematic drawing of screw (blue) and edge (red) dislocations on 111 glide-plane within Si. g Corresponding zone-axis diffraction pattern of c-Si phase in e. h 4D-TEM is carried out to generate elastic strain maps with ϵxx, ϵyy and Θ demonstrating a significant increase in strain close to the interface of the Si-domain. Strain regions correlate with observed areas indicating high density of structural defects.

Full size image

Figure 4d presents a DF-STEM image of a silicon domain after three cycles, slightly deviated from the [100] zone axis, providing deeper insights into the initial structural deformation of the crystalline Si-domain near its interface at low cycling number. After just a few lithiation cycles, a notable density of crystallographic defects or dislocations are apparent, giving rise to the formation of amorphous shear bands within the c-Si interface region. The grown SEC at the Si-interface is highlighted, corroborated by the EDS data in Supplementary Figure 12. Additionally, a pore is indicated towards the lower right. A more detailed analysis of the emerging dislocations upon electrochemical cycling, is presented in Fig. 4e. The DF-STEM image provides a magnified view of a silicon domain adjacent to the domain shown in Fig. 4d (see Supplementary Figure 13). Dislocations appear as narrow lines, characterized by a Burgers vector b = <110> and slip plane {111}46. Their slip trace fits well with a {111} plane. Possible dislocations for screw (blue) and edge (red) dislocations on a {111} slip plane (grey) are depicted in a schematic unit cell drawing, see in Fig. 4f. Approximately 8 screw and 9 edge dislocations are identified in the vicinity of the Si-domain interface within the section after the third cycle. Figure 4g shows the corresponding zone-axis diffraction pattern for the c-Si phase. The observations suggest that the dislocations emerge and their density gradually increases with electrochemical cycling. The volume swing triggered by ion insertion and extraction generates alternating mechanical loading conditions and non-uniform strain, leading to crystallographic defects within the silicon domain as depicted in Fig. 4e. Finally, amorphous shear bands form throughout the entire silicon domain, see Fig. 4a. The resulting phase transformation of the crystalline silicon domain affects the lithiation process at its interfaces, as illustrated by the modified interface microstructure after prolonged cycling in Fig. 2c, e. Depending on the phase of the Si-interfaces, stronger or weaker lithiation occurs, visualized by the non-uniform dendrite growth, which in turn affects the strain distribution acting on the silicon and ultimately the capacity.

To quantify the deformation-induced strain, elastic strain maps ϵxx and ϵyy by utilizing 4D-STEM are generated after three electrochemical cycles, see Fig. 4h. These maps reveal intricate patterns of multiaxial tensile and compressive strain concentrations, with magnitudes reaching up to ±2%. Additionally, a rotation Θ of the principal strain direction, reaching up to 2° within the Si domains, is observed. In particular strong strain variations are visible at the interface of the Si domain. The maps indirectly confirm a significant deformation in the near-surface regions of the Si domains, attributed to volume changes during electrochemical cycling. Notably, high strain is detected, particularly in proximity to domains associated with dislocations. It is important to highlight that the strain map was acquired after the third delithiation. Strain analysis of the sample after 300 cycles was unfeasible due to exceedingly high strain rates, precluding the derivation of meaningful results.

Mechanical loading and Si-particle phase transformation

Molecular dynamics (MD) simulations, see Methods, are used to validate empirical observations regarding the emergence, branching, and interaction patterns of shear bands, as well as the stabilization and widening of dominant shear bands caused by alternating compression and tension load. The presented computational approach is designed to offer a purely mechanical perspective, replicating the insertion and extraction of Li-ions and the ensuing volume swing of the Si particle, as illustrated in Fig. 5a. The schematic illustrates the principle of lithiation and delithiation of a crystalline silicon domain. Upon lithiation, the interaction of Li-ions (blue) and Si-atoms (red) results in the formation of a high-lithium-content LixSiy alloy, enveloping the non-lithiated crystalline silicon core. Simultaneously, this process leads to the formation of the SEI, highlighted in green. During lithiation, the silicon core experiences compression as the volumetric expansion of the lithiated shell, indicated by black displacement \({u}^{ \rightharpoonup }\) arrow, is confined by the surrounding matrix, generating internal stresses within the particles. Conversely, delithiation causes a reduction in the volume of previously lithiated regions, alleviating again the pressure on the silicon particle core. The core region experiences non-hydrostatic strain due to the non-uniformity of lithiation across the particle surface. In the course of cyclisation, the SEI layer repeatedly breaks up and new SEI is formed, leading to the gradual local thickening and SEC layer formation at the interface of the silicon. Overall, compressive stresses will act on the remaining crystalline silicon core and additionally stress accumulates on the surrounding SEC and the CBD. This is in accordance with the experimental data presented in Fig. 4h where the elastic strain maps εxx and εyy show significant tensile distribution on the interface of the silicon domain. The inset in Fig. 5a on the right, schematically illustrates the displacement experienced by the silicon core of the particle during electrochemical cycling, utilized for the MD simulations. The process begins with a compressive length change, followed by an idealized dilation step back to the initial extend of the core due to delithiation. Within the modelling, a single crystalline silicon domain oriented along the [100] direction experiences alternating loading conditions utilizing a plane stress state. This approach ensures a well-defined maximum shear stress oriented at a 45-degree angle to the loading direction, avoiding complex stress states and ensuring clear directionality for stress components, see Supplementary Note 6 for detailed methodology as well as Supplementary Figure 14.

Fig. 5: Molecular dynamics simulation to reproduce the mechanical perspective of deformation-induced phase transformations.
figure 5

a Schematic (not to scale) demonstrates the volume swing experienced by the Si particle during the insertion and extraction of Li+-ions, emphasizing the corresponding expansion and shrinkage over time. This dynamic process is highlighted by bold black arrows indicating the displacement \({u}^{ \rightharpoonup }\) of Si atoms for lithiation and delithiation steps, respectively. During lithiation, the silicon atoms towards the core experience compression, whereas during delithiation, they experience alleviation. The MD simulation inset replicates these electrochemical cycling effects on the silicon core through alternating compression and tension conditions, which are subsequently utilized for the MD modelling. b MD modelled time series illustrating the formation of shear bands under mechanical loading. The emerging shear band is outlined by dashed lines, with distinct phases represented by colours: diamond-cubic (dc)- Si in red, diamond-hexagonal (dh)- Si in orange, a-Si in yellow and the strongly distorted dc-Si phase in blue. c MD modelling for a larger sample. d BF-STEM image highlights the formation of broad shear bands after 300 electrochemical cycles on a similar length scales as for the MD modelling.

Full size image

Figure 5b displays a time series of the sample’s cross-section, with individual timesteps from I to VII according to the schematic presented in Fig. 5a. The model provides information with respect to different phases comprising diamond-cubic silicon (dc-Si), amorphous Si (a-Si), diamond-hexagonal silicon (dh-Si) and a strongly distorted diamond-cubic Si phase, highlighted in red, yellow, orange and blue, respectively. In its pristine, unloaded state (I), the sample only exhibits artificially created voids. The initiation of compressive loading (II and III) induces dislocation emission, accompanied by trailing stacking faults along various {111} planes. These stacking faults promptly transform into a shear band as the dislocations advance under the applied load. All shear bands align with the direction of maximum shear stress at approximately a 45-degree angle. The smaller amorphous regions located at the figures’ centre correspond to intersections where propagating shear bands intersect at cross-sectional slices. Within the interface of a-Si and dc-Si, a thin layer of distorted dc-Si is noticeable. Subsequent tensile loading (IV) induces further amorphization as shear bands propagate in response to changing load conditions. Additionally, strips of diamond-hexagonal silicon (dh-Si) are observed in this region. It is noteworthy that the molecular dynamics (MD) simulation captures transitions from dc-Si to a-Si, as well as from dc-Si to dh-Si followed by a transition to a-Si. See also Fig. 3b, c in this context. Furthermore, shear branching occurs due to local stress state alterations caused by plastic deformations of surrounding areas. These processes continue in the second compressive cycle. After the second tensile loading (VI), pre-existing branched shear bands consolidate, and remaining stacking faults undergo complete amorphization, leading to the growth of shear bands across the entire cross-section. This results in the emergence of a dominant shear band, serving as the primary site for plastic deformation and thickening to a few nanometres. Subsequently, the sample predominantly elongates due to shearing in the amorphous shear band. This elongation is facilitated by the gliding motion of the upper and lower crystalline halves against each other, causing further thickening of the shear band, primarily through smaller bands branching at the crystalline-amorphous interface. The variation in the appearance of smaller bands within different figures arises from the gliding motion occurring on a slightly tilted (111) plane, revealing different atoms in the cross-section. However, away from the shear band, regions remain relatively stable, undergoing minimal change after the evolution of the dominant shear band. After the last step (VII), marginal differences are observed compared to the preceding step. Furthermore, despite multiple additional pure compression and tensile deformation cycles, the thickness of the shear band no longer increases.

Following the in-depth analysis of a specifically designed sample capturing atomistic mechanisms driving amorphization, additional simulations on a larger sample are performed by varying not only the loading conditions but also the loading direction, see Fig. 5c. Under the more complex conditions, this simulation reveals a single crystalline matrix with interspersed amorphous bands between the crystals. Notably, regions where shear bands interact have larger amorphous regions, indicating further shear band branching. Importantly, this final microstructure closely aligns with the experimental observations illustrated in Fig. 5d, wherein interface regions between c-Si and a-Si exhibit transition zones, likely linked to the distorted dc-Si phase observed in the simulation. In the corresponding HR-TEM image, the zone axis is [\(\bar{1}14\)]. The direction of the observed interface suggests that the interface between the amorphous and crystalline phase is a (\(1\bar{1}\bar{1}\)) plane, which fits well to the interface found in MD simulations. The formed amorphous shear bands in contact with the Si interface at prolonged cycles will influence the lithiation process, with more lithiation occurring at the amorphous phase. Consequently, a non-uniform SEC growth on the silicon interface is observed, see Fig. 2c, e.

Discussion

In this study, we provide crucial evidence emphasizing the importance of considering amorphous shear band formation within crystalline silicon, in addition to the conventional understanding that mainly attributes capacity fading in silicon anode LIBs to SEI reformation upon lithiation and delithiation. Based on the structure-property relationship, a qualitative correlation between microstructural modifications during electrochemical cycling and capacity fading is made. Employing a multi-scale and multi-method approach, we elucidate structural and chemical changes while considering the varying mechanical conditions experienced in silicon-rich anodes. We commence with a thorough investigation of structural changes on large scales, obtaining crucial statistical information using nano-tomography and ML-based image analysis. In combination with a detailed analysis on small scales using FESEM and TEM techniques we robustly confirm significant growth of the SEC during cycling, alongside a reduction in the size of silicon particles and pore space. While pore space may partially alleviate the expansion of the active material by acting as a stress buffer during lithiation, the rigidity of the SEI leads to its frequent breakdown and reformation on the silicon surface. The continuous SEI reformation and SEC growth as illustrated by the comprehensive ML-based microstructure analysis significantly degrades the silicon content as well as depletes usable lithium ions and electrolyte25, thereby playing a substantial role in the observed capacity fade through inducing substantial loss of active material.

Further, our observations uncover a non-uniform growth of the SEC on the silicon domains interfaces, which likely affects the capacity behaviour, see in this context Fig. 2c. and Fig. 1d, respectively. Building upon prior hypotheses suggesting that particles with smooth surfaces and no structural alterations were either never electrically connected or were disconnected during cycling10, or influenced by the homogeneity of the microstructure27, we propose that the non-uniform SEC growth is in addition a consequence of uneven Li-ion flux through the SEC45.

Literature11,14,20,41,47,48,49,50 indicates varying lithium diffusion rates across different crystal facets of Si, with the <110> direction being more thermodynamically favourable due to atomic channels along this direction in the diamond cubic crystalline structure. The crystallographic anisotropy causes a predominant volume expansion along the {100} surfaces of silicon, which affects SEI breakup and subsequently triggers anisotropic SEC growth. Lithiation depends on the underlaying phases, with a higher lithiation speed for amorphous Si, consistent with recent research5,51. Our findings support the prevailing view of amorphization at the silicon interface, which entails the formation of a core-shell structure characterized by a crystalline silicon domain constituting the core and the amorphous LixSiy forming the shell, as reported in literature21. As the electrolyte permeates along percolation channels formed by evolving voids during lithiation/delithiation cycles, the SEI forms on the surface of these voids and progressively grows inward. This is substantiated by the presented STEM investigations, see Fig. 2e, which reveal plate-like growth stages with a thickness of approximately 23 nm, supporting the process of SEI reformation. This reinforces the general conjecture attributing battery degradation to the loss of active materials, such as Li and Si, during SEC formation (Fig. 1c, Supplementary Figures. 6 and 7) and the abatement of the Li-diffusion capabilities towards the active material.

Note that the electrolyte salt impacts the SEI formation52 and thus influences the SEC growth. A systematic investigation of the effect of various salts or electrolytes on the SEC growth is beyond the scope of this work; however, it is definitely an area worthy of further thorough investigation.

In addition to the previously reported induced amorphization linked to the SEC formation10,21,23,24,25 at the silicon interface, this study identifies a stress induced transformation from crystalline to amorphous within the silicon particle’s core. The SEC growth and phase-dependent lithiation speed causes non-hydrostatic strain in the silicon core. This occurs as the core is compressed as the expanding lithiated shell is confined by the surrounding matrix, creating internal stresses. Our experimental findings, in alignment with MD simulations, reveal dislocation formation in the samples after three cycles, particularly at the interface between the SEC and the c-Si phase, see Fig. 4d, e. As such, the MD simulations depict a purely mechanical perspective on cycling, highlighting the importance of volumetric changes in the Si core in the occurrence of amorphous shear bands. It is important to acknowledge that our MD simulation is simplified, comprising of controlled constant strain-rate compression and dilation. Therefore, it has implications for subsequent lithiation and delithiation steps, potentially influencing shear band formation and introducing deviations between the simulation and real-world behaviour. Nevertheless, it provides qualitative insights into the underlying mechanism, as highlighted by the alignment between experimental and model results, see Fig. 5c, d. The MD simulation shows that dislocations accumulate in localized deformation zones, forming amorphous shear bands that traverse the silicon core. After 300 cycles, broad bands of amorphous silicon become evident, penetrating the crystal as a consequence of accumulated damage, see Fig. 4a, b. This observation aligns with prior investigations documenting solid-state amorphization phenomena in crystalline silicon subjected to external pressure30,31,53. He et al.31 conducted experiments applying compressive stress on a silicon nanopillar, revealing the formation of a localized band characterized by perfect dislocations. Within this band, a newly identified phase layer with a diamond-hexagonal (dh-Si) structure was identified, followed by the generation of amorphous nanodomains upon subsequent compression. Notably, He et al. reported dislocation nucleation occurring at a shear stress of approximately 6 GPa, with MD simulations estimating the onset of dislocation nucleation to be around 1–2 GPa. Similarly, Zhao et al.30 observed the penetration of amorphous bands into the crystal under pressure through laser shock experiments conducted on millimetre-sized silicon cylinders, with their findings further supported by MD simulations.

Studies investigating the lithiation of crystalline silicon surfaces35,51,54,55,56,57, have reported deformation and induced stresses in LixSi reaching to several GPa. Plastic deformation of silicon electrodes during electrochemical lithiation has been reported at compressive stress levels of around 1.75 GPa55. Particularly, high compressive stress can accumulate at the Li-Si phase boundary of crystalline silicon during lithiation, with values reaching approximately up to 6 GPa51. Hence, we argue that the previously reported stress values are sufficient to initiate plastic yielding and shear band formation upon lithiation in Si-based cells, as discussed here. Moreover, the observed formation of the shear bands within the silicon core domains provides an estimate about the underlying stresses present within Si-anodes, with values in the GPa regime. Yet, we point out that the ensemble effects and a quantification of the stresses across the investigated electrode remain unknown.

Further, we would like to point out that while silicon may experience ultimate failure in a brittle manner58,59, the silicon domains embedded in the CBD may also benefit from the observed solid-state amorphization of the silicon core. Research on crystalline metals has demonstrated that crystalline-amorphous nanostructures can exhibit enhanced strength and improved stability in plastic flow60. In this context, amorphous domains play a role in impeding dislocation motion. Additionally, crystalline-amorphous interfaces demonstrate unique inelastic shear transfer properties, effectively inhibiting fracture initiation. Amorphous silicon exhibits lower hardness and elastic modulus, thereby displaying higher fracture toughness compared to crystalline silicon51,61. This is likely attributed to its superior capability to deform and mitigate stress. The disordered structure and reduced atomic interactions in amorphous Si facilitate stress relief through deformation. Given its lower strength compared to crystalline Si, induced stress in a-Si can be alleviated through the rearrangement of Si atoms62. Our findings, in line with a recent investigation in crystalline materials under mechanical loading58, suggest that the small volume change during the phase transformation from c-Si to a-Si allows shear bands to drive plasticity and ductility rather than serving as precursors to fracture, suppressing void nucleation. The energetically favourable formation of amorphous shear bands, as opposed to cleavage, aligns with our investigations that did not identify noticeable fractured silicon particles, in particular at prolonged cycling. This challenges reports suggesting Si anode failure primarily occurs through fracture. While our experimental observations are limited to samples in the delithiated state, thus lacking insights into the anode’s lithiated state, both experimental investigations and MD simulations show no initiation of crack formation. This supports our assertion that the formation of thin amorphous regions within the crystalline silicon core could potentially impede complete particle pulverisation.

In conclusion, our study sheds light on the continuous capacity loss observed in LIBs employing silicon-rich anodes throughout extended cycling without instantaneous failure within 300 cycles. It reveals the complexity of degradation mechanisms in silicon-based anodes, including the SEC growth and phase-dependent lithiation speed. Both generate non-hydrostatic strain in the silicon core, triggering dislocation accumulation and finally deformation induced amorphization within the silicon core domain. The shear band formation boosts non-uniform growth of the SEC on the silicon interface, contributing to the overall stress configuration within the anode and affecting battery cycling behaviour. The mechanistic insight underscores the importance of delving beyond surface chemical processes, such as SEI formation, to grasp the complexities affecting the entire battery system. The insights derived from smaller-scale investigations effectively elucidate the failure mechanisms in realistic, large-scale electrodes utilizing crystalline Si particles. Furthermore, we argue that while the deformation triggered phase transition impacts the lithiation and SEC formation process, over prolonged cycles, it could potentially yield beneficial effects by inhibiting fracture initiation and silicon particle pulverization slowing down the total battery failure. Overall, it is imperative to emphasize the significance of the homogeneity of the initial lithiation, as it plays a pivotal role in dictating overall capacity behaviour.

The provided understanding extends beyond the current study to encompass a wide range of silicon-based electrode materials, including emerging technologies like all-solid-state-batteries. By comprehending the full spectrum of mechanisms at play during battery operation, we hold promise for averting issues like particle pulverization and catastrophic cell failure, thus paving the way for future advancements in battery technology.

Methods

Electrode preparation

The electrode paste was produced using a Hivis Mix Modell 2P-03/1 mixer. The paste was subsequently coated onto a 10 µm copper foil from Schlenk. For the electrode paste a composite mixture of 70 wt% crystalline silicon particles (Silgrain® e-Si 408), 22.8 wt% graphite particles (BTR S360 L2-H), and 0.2 wt% carbon nanotubes (TUBALLTM BATT H20) was used, all interlinked with 5 wt% sodium carboxymethyl cellulose (Na-CMC) and 3 wt% styrene-butadiene rubber (SBR). These electrodes exhibited a mass loading of 3.8 mg cm−2 and an area capacity of 3.6 mAh cm−2.

Cell fabrication and electrochemical cycling

For the cell assembly and electrochemical cycling, pouch cells were fabricated in a full-cell configuration, employing the prepared negative electrode, a polyolefin separator, and a nickel–manganese–cobalt-oxide (NMC622 2.8 mAh cm−2) cathode coated onto an aluminium foil. A schematic representation can be found in Fig. 1a. These pouch cells were then enclosed within laminated film pouches and filled with 300 µl of FEC/EMC 1:2, 1 M LiPF6 + 2 wt% VC electrolyte before being sealed, all conducted within an argon-filled glove box. The cells were then cycled with a charge/discharge rate (C-rate) of 0.5 in the voltage range of 3.0 V and 4.3 V, with a constant voltage charging step at 4.3 V for 1 h, using a Maccor battery tester. Lithiation of the anode was conducted until reaching a specific capacity of 1200 mAh g−1. After 24 cycles, cycle 124, cycle 224 and cycle 324, a rate capability check was performed with 0.2 C, 1 C, 2 C and 3 C.

After the 3rd discharge step, one cell was opened within an Ar filled glovebox. The anode was extracted from the cell and left in the glovebox for 1 h to facilitate the evaporation of the electrolyte. A similar procedure was followed after the 300th discharge step to obtain samples representing the 3rd and 300th discharge steps.

Synchrotron X-ray nano-tomography

X-ray nano-tomography scans were carried out at the ID16B beamline40 of the European Synchrotron Radiation Facility (ESRF) in Grenoble, France, using holo-tomography63 at voxel sizes of 25 nm, 50 nm and 100 nm. An holo-tomographic scan consists of performing phase contrast 3D imaging at four different distances between the detector and the sample to slightly change the propagations distances. A conic and monochromatic beam with an energy of 29.6 keV and a flux of 1012 ph s−1 was used. 2505 projections, as well as 20 and 21 reference and dark images were recorded for each tomographic scan on a pco.edge camera (2048 × 2048 pixels) with an LSO30 scintillator along a 360° rotation with an exposure time of 20 ms per step. The total acquisition time was approximately 10 min per holo-tomography scan.

The 3D reconstruction process entailed an initial phase retrieval calculation, executed through a custom in-house octave script based on a Paganin-like approach using a delta/beta ratio of 170. Subsequently, a filtered backprojection reconstruction was performed using the ESRF software PyHST264. An in-house post-processing script was applied on the reconstructed data to reduce ring artefacts. Final reconstructed volumes of 51.2 × 51.2 × 51.2 µm³, 102.4 × 102.4 × 102.4 µm³, and 204.8 × 204.8 × 204.8 µm³, corresponding to voxel sizes of 25 nm, 50 nm, and 100 nm, respectively were obtained in a 16-bit unsigned integer data range format.

Machine-Learning image segmentation

To segment the different phases in the reconstructed nano-tomography volumes, a convolutional neural network approach was applied. U-Net models from the Python Keras library were employed, which were trained with datasets comprising several images and their respective labels distributed throughout the entire reconstructed volume. These labels were generated through an initial grey value threshold segmentation using Python, followed by manual refinements using the open-source software, Ilastik65. The models were trained on images of size 2048 × 2048 pixels for 100 epochs on an NVIDIA Quadro RTX5000 GPU. For more details see Supplementary Note 2. Following the machine learning-based prediction, the segmented data underwent post-processing using Python libraries including scikit-image, numpy and scipy. The final rendering and 3D visualization of the processed data was accomplished using Avizo®3D (version 2022.2).

FESEM measurement

A Hitachi IM4000+ ion milling system was used for the preparation of the anode cross sections. The system uses low-energy Ar ions to perform cross section cutting without applying mechanical stress to the sample. The FESEM investigation was carried out using a ZEISS GeminiSEM 450 using an acceleration voltage of 5 kV and a current of 3 nA. Imaging was done at a magnification range between 100 and 30.000, achieving 5.58 nm pixel size and 18.61 nm pixel size at a magnification of 10 kx and 3 kx, respectively. The EDS measurements were conducted using an Oxford Ultim Extreme detector with 1024 × 768 pixels using the voltage and current settings of 2/3 kV and 3 nA, respectively.

FESEM-EDS data processing

Similar to the X-ray nano-tomography data processing, a machine learning-based segmentation approach was used to develop predictive models for segmenting the FESEM images (refer to Supplementary Note 4). This ML-based approach presented a significant advantage over conventional segmentation methods by effectively distinguishing different phases, even when they exhibited similar grey values (e.g., distinguishing background regions). To achieve accurate segmentation, two distinct models were employed: one tailored for pristine, one for samples after 3-cycles and another specifically designed for samples subjected to 300 cycles. The segmented images were subsequently utilised to extract valuable statistical information from the samples, including particle size and phase volume evolution. Additionally, the labelled image derived from segmentation served as a mask for the EDS mappings, facilitating precise calculations of element occurrences while avoiding potential misinterpretations of detected information within pore regions.

Transmission Kikuchi diffraction analysis

Electron backscatter diffraction (EBSD) was carried out in transmission mode using a Zeiss GeminiSEM 450 in combination with an Oxford symmetry 2 detector with 1244 × 1024 pixels. Lamellas of the pristine anode, the anode after three cycles and after 300 cycles were first prepared using a Zeiss Auriga 40 Crossbeam. For the TKD measurement, voltage and current were set to 30 kV and 20 nA, respectively. A pixel size of 5 nm, 3 nm and 5 nm was achieved for the pristine, 3-cycled and 300-cycled sample, respectively. The samples were placed on a 20° pre-tilted sample holder. The software AztecCrystal by Oxford instruments was used to process and analyse the acquired images.

HR-STEM and 4D-STEM

Transmission electron microscopy (TEM) was carried out using a JEOL 2200FS system operating at 200 kV. Bright-field TEM and scanning TEM (STEM) were used to image the structure at the nanoscale. Dark-field TEM imaging was used for dislocation characterization and to determine the crystal structure, selected area diffraction (SAD) patterns were recorded using a selected area aperture corresponding to a diameter of 200 nm. In addition, to obtain the local atomic structure at the nanoscale, high-resolution TEM (HR-TEM) was recorded. By computing the FFT from subsets of the HR-TEM images, local diffractograms were obtained. Finally, 4D-STEM was used to acquire maps of electron nanodiffraction patterns. The scanning system was controlled using a TVIPS universal scan generator and the diffraction patterns were recorded with a QuantumDetector Merlin4R direct electron detector. A small condenser aperture was used to obtain a small rather parallel probe (~2.5 mrad convergence angle). From the nanodiffraction patterns elastic strain maps were computed using a custom code66. The positions of the diffracted peaks were detected using template matching and reference region was defined in a dislocation free area, allowing to calculate the strain tensor at each probe position.

Simulation

Cyclic mechanical testing and preceding sample preparation was carried out with the large-scale molecular dynamics simulation software LAMMPS67. The simulation system comprised two samples depicted in Fig. 5b, c, respectively: a smaller sample (Fig. 5b) containing 580,000 atoms and a larger (Fig. 5c) one with 5 million silicon atoms. These samples were subjected to periodic boundary conditions within simulation cells measuring 55 × 21 × 11 nm and 130 × 87 × 119 nm, respectively. It is important to note that the MD simulation may not capture the long-time scale of the experiment, potentially resulting in structural deviations. The atomic interactions were described using the Stillinger–Weber potential68, known to qualitatively well reproduce the behaviour of silicon69 while maintaining computational efficiency. The simulations were conducted within a constant number, pressure, and temperature (NPT) ensemble, maintaining a temperature of 300 K. The Velocity-Verlet algorithm was employed with a time step of 2 fs. When pressure control was necessary in a specific direction, it was set to 0 GPa using a Nose–Hoover barostat. To establish a well-defined stress state and deformation site, the larger sample was pre-notched to mimic the effect of the surrounding phases on the Si particle. The size and geometry of these notches were selected to produce a dominant shear band rather than multiple small shear bands, which result in an overly refined microstructure. The testing protocols encompassed cyclic uniaxial loading and unloading at a constant strain rate for the smaller sample, while the larger sample underwent direction-alternating uniaxial loading and unloading under the same conditions. Total deformation of approximately 20% of the loaded dimension was used. Each cycle alternated between compression and extension along both the x-axis and y-axis for the larger sample. The duration of each un-/loading run was 0.8 ns (4,000,000 simulation steps) and 0.2 ns (1,000,000 simulation steps) for the small and large sample, respectively. The results of the MD simulations were analysed with OVITO70.