## Introduction

Solar cells based on organic–inorganic hybrid perovskites (CH3NH3PbI3, MAPbI3) have attracted widespread research attention due to their low synthesis cost and high-power conversion efficiency (PCE)1,2,3. Despite the great progress in increasing the device PCE from initial 3.8%4 to the most recent 23.3%5 in just one decade, the commercialization of this technology remains hindered by the stability issues of materials. It has been reported that the perovskites rapidly degrade under increased temperature6, oxygen7, moisture8, and UV light illumination9, often due to the structure instability caused by electromigrations10, ion migration,11 and interfacial relationships12. In fact, with increasing temperature, visible diffusion of iodine initiates at a temperature below 150 °C and lead migration is induced at a higher temperature around 175 °C, causing the degradation of perovskite and formation of PbI26. The ion migration induced by the electric field, heat, and light illumination also contributes to the hysteresis in J–V curves13,14, which in turn leads to poor long-term stability resulted from significant structural changes, including the lattice distortion and the phase decomposition. Thus, it is essential to characterize the intrinsic structure as well as structure stability and the decomposition pathway of these materials to improve organic–inorganic hybrid perovskites, which motivates our present transmission electron microscopy (TEM) study on perovskite MAPbI3.

Although TEM has been widely used for structure characterization of perovskite MAPbI315,16,17, the atomically resolved imaging remains elusive due to its electron beam-sensitive nature18. Thus, the electron diffraction (ED) or fast Fourier transform (FFT) pattern based on lattice fringes becomes a common preference to identify the phase of perovskite. However, in most of the experiments, the observed ED or FFT patterns do not really match those of a perfect tetragonal perovskite phase exactly16,17,19,20,21,22,23,. For example, {110} diffraction spots along the [001] direction are usually not observed in electron microscopy even though these reflections are visible from X-ray diffraction (XRD) measurements24,25,26,27,28,29. Long et al.24 identified a textured thin MAPbI3 film to be a tetragonal [001] zone axis even though {110} spots were not observed. Park et al.25 prepared a cross-sectional whole device of glass/FTO/mp-TiO2/MAPbI3/spiro-OMeTAD, and similarly the {110} spots were absent. Polarz et al.26 developed metal–organic porous single-crystalline MAPbI3 and also did not observe the (1$$\bar 1$$0) spots. Furthermore, Tang et al.27, Vela et al.28, and Zhu et al.29 found that the {110} spots were missing in MAPbI3 nanowires. Segawa30 explained that the (110) and (1$$\bar 1$$0) diffraction spots are inherently difficult to be detected experimentally due to the low ED intensity of organic–inorganic halide compounds. Despite the absence of some expected diffraction spots, the corresponding structures were often identified as a tetragonal perovskite phase, as summarized in Supplementary Table 1. Nevertheless, {110} diffraction spots of perovskite31 have been captured from a rapid acquisition with the dose rate as low as 1 e Å−2 s−1, suggesting that the absence of {110} reflections of perovskite may be related to structure degradation during electron microscopy characterizations. From an energetic point of view, electron beam illumination, optical illumination, heating, and the electric field are similar, enabling ions to overcome the migration barriers and thus inducing the structure transition32,33,34. Revealing the structure evolution of MAPbI3 under electron illumination, especially its decomposition pathway, therefore, can help us to understand material degradation in practical solar cell devices.

In this paper, we use atomically resolved Z-contrast imaging, diffraction analysis, and quantitative energy- dispersive X-ray spectroscopy (EDX) in an aberration-corrected TEM in an attempt to understand the structure and structure evolution of single-crystalline tetragonal MAPbI3. We found that the perovskite structure undergoes complex phase changes under the electron beam illumination. The decomposition of MAPbI3 is likely initiated with the loss of I ions to subsequently form superstructure MAPbI2.5, followed by the loss of MA and additional I ions to form MAyPbI2.5–z (0 ≤ y ≤ 1 and 0 ≤ z ≤ 0.5) accompanied with the disappearance of a superstructure, leading to its final decomposition into PbI2. During this process, the ED patterns change subtly that is often overlooked, and thus the evolving and decomposed phases are often incorrectly identified as perovskite. Indeed, if we can only observe {2h, 2k, 0} diffraction spots along the [001] direction and while the {2h+1, 2k+1, 0} reflections [e.g., (110)] are absent, it is highly likely that the tetragonal perovskite phase has already decomposed into a hexagonal PbI2 structure along the [$$\bar 4$$41] zone axis. We also identify the critical electron dose limit to obtain the intrinsic perovskite structure. Our findings impose important questions on some earlier experimental interpretations of ED measurements of MAPbI3 and highlight the need to circumvent material degrading in future experiments. The identified optimal TEM conditions to obtain the intrinsic perovskite structure can be used to guide the future electron microscopy characterization of these materials. The understanding of the degradation process also sheds light into the general stability issue of the organic–inorganic hybrid perovskites in solar cell applications.

## Results

### Mismatches in ED patterns

MAPbI3 crystals were self-grown on FTO/TiO2 substrates, as reported in our previous study35,36. The powder XRD (Supplementary Fig. 1a) indicates a good crystallinity of tetragonal MAPbI337,38. The single-crystalline nature is further confirmed by synchrotron X-ray diffraction (XRD) (Supplementary Fig. 1b) with the measured (004)c reflection denoted in the pseudo-cubic setting. However, when MAPbI3 crystal was put into TEM, the acquired ED and FFT patterns do not match those of a perfect tetragonal perovskite structure. The atomistic models of tetragonal MAPbI3 along [100] and [001] axis are shown in Fig. 1a, b, and the corresponding ED patterns simulated from the perfect structure are shown in Fig. 1c, e, and g. A subtle difference between the experimental SAED (Fig. 1d and h) and FFT pattern (Fig. 1f) acquired from regions in Supplementary Fig. 2 and the simulated ones can be noted after careful comparison. For example, Fig. 1d is identified to be the [001] zone axis of tetragonal MAPbI3, which is similar to the simulated ED pattern in Fig. 1c, yet {110} diffraction spots do become dimmer. More importantly, Fig. 1f is identified to be the [210] zone axis, wherein the (002) diffraction spot can hardly be seen. Figure 1h is identified to be the [10$$\bar 1$$] zone axis, wherein the (020) spot is not observed while an additional (101) spot appears. Therefore, whether these SAED patterns indeed correspond to tetragonal MAPbI3 is questionable and needs further investigation.

### Atomic structure of PbI2

To resolve the subtle difference between experimental and simulated ED patterns, atomically resolved Z-contrast images were recorded at 80 kV. An atomically resolved STEM image in Fig. 2a shows stripes with a measured distance of 0.72 nm. Based on the corresponding FFT pattern (Fig. 2b), the viewing direction is often mistaken as the [10$$\bar 1$$] of MAPbI3 without noticing that additional (101) reflection appears while the (020) spot is missing. After careful analysis, we found that these stripes are in fact not from MAPbI3 but PbI2, which can be illustrated by the ball-and-stick model of the PbI2 (Fig. 2c). Furthermore, simulated HAADF-STEM (Fig. 2e) was carried out to compare with the experimental image contrast (Fig. 2a), which shows good agreement. The experimental FFT pattern (Fig. 2b) can be identified to be the [120] direction of PbI2, which matches the simulated ED pattern (Fig. 2d) well. The detailed atomistic configurations of the hexagonal PbI2 (space group: R-3m:H, a = b = 0.4557 nm, c = 2.0937 nm, α = β = 90°, and γ = 120°) are shown in Supplementary Fig. 339, wherein the layered PbI2 structure is composed of Pb–I octahedra. The stability of this structure at room temperature is further confirmed by the molecular dynamics (MD) simulation at 300 K.

Additional atomic structure data along different directions (Fig. 2f–t) further confirm that the original sample of MAPbI3 is indeed decomposed into PbI2. These STEM images have been filtered and the pristine images are shown in Supplementary Fig. 4. Figure 2f shows the alternating Pb and I arrangements of PbI2, illustrated by the ball-and-stick model in Fig. 2h. The FFT pattern (Fig. 2g) from the region (Supplementary Fig. 4b) is identified to be the [8 10 1] zone axis, which is consistent with simulated ED pattern (Fig. 2i). In addition, the simulated STEM image in Fig. 2j also shows a similar contrast, compared with Fig. 2f. Likewise, atomic STEM images along [$$\bar 1$$11] and [110] zone axes have also been analyzed as shown in Fig. 2k–t. These STEM images well match the corresponding ball-and-stick models and simulated STEM images of the PbI2, and the corresponding FFT patterns are also consistent with the simulated ED patterns of the PbI2. Additional data of PbI2 are presented in Supplementary Fig. 5, showing a subtle difference with simulated MAPbI3 in the ED pattern in Supplementary Fig. 6. It is worth noting that both ED patterns of [8 10 1] and [$$\bar 1$$11] zone axes of PbI2 are often easily mistaken for the [10$$\bar 1$$] direction of MAPbI3 when the missing of (020) diffraction spot from MAPbI3 is not recognized.

### Structural comparison between MAPbI3 and PbI2

It is quite revealing to highlight the difference in atomic structures and ED patterns between MAPbI3 and PbI2. Figure 3a is a HAADF-STEM image of PbI2 viewing along the [$$\bar 4$$41] direction judging from the corresponding FFT pattern (Fig. 3b) as well as the SAED and TEM images (Supplementary Fig. 7). The atomistic model, simulated ED pattern, and simulated STEM image of PbI2 along this zone axis are shown in Fig. 3c–e. However, in many previous studies16,17,19,20,21,22,23,24,25,26,27,28,29, this ED or FFT pattern has been mistakenly identified as [001] zone axis of tetragonal MAPbI3 without noticing the absence of {110} diffraction spots (Supplementary Table 1). For comparison, based on the atomistic model in Fig. 3f, we also carried out the simulations of ED (Fig. 3g) and HAADF image (Fig. 3h) for MAPbI3 along the [001] zone axis. Note that the distance in reciprocal space for (220) reflection of tetragonal MAPbI3 is 0.3196 Å−1 (Fig. 3g), which is extremely close to that of (104) and (0$$\bar 1$$4) reflections (0.3173 Å−1) for PbI2 (Fig. 3d), making it difficult to distinguish them only based on the SAED pattern. Despite of the similarity in their ED patterns, the simulated atomic structures of these two phases are quite different. For tetragonal MAPbI3 along the [001] zone axis, the contrast of atom columns is not uniform (Fig. 3h), in comparison with the uniform contrast in PbI2 along [$$\bar 4$$41] in Fig. 3e. Moreover, the distance between two brighter atom columns is 0.625 nm for MAPbI3 (Fig. 3h), which is twice the distance between two brighter atom columns (0.315 nm) of PbI2 (Fig. 3e).

### The structure evolution during decomposition

The evolution of tetragonal MAPbI3 decomposing into hexagonal PbI2 is important for understanding the material degradation, and it can be captured at a low magnification with SEM and a series of EDX mappings (Fig. 4). During the process, the I/Pb atom ratio decreases from the initial 2.62 to the final 1.92 due to the volatilization of the iodine40,41 (Supplementary Fig. 8), consistent with the formation of PbI2. The quantitative EDX mapping in the STEM mode at 80 kV also shows that the average atom ratio for I/Pb is close to 2 (Supplementary Table 2, based on the spectrum in Supplementary Fig. 9). Additional STEM-EDX mapping was conducted on other samples (Supplementary Fig. 10) showing similar observations.

In order to capture the intrinsic perovskite structure and investigate the detailed degradation process, the evolution of ED patterns under a low-dose rate ~1 e Å−2 s−1 is recorded in Fig. 5a–d viewing along [001] MAyPbI3–x (0 ≤ y ≤ 1 and 0 ≤ x ≤ 1) (or $$[\bar 441]$$ of PbI2). These time-series SAED patterns match the corresponding simulated ones (Fig. 5e–h) based on the atomistic structures (Fig. 5i–l). During the decomposition, we observe an intermediate phase with superstructure spots (Fig. 5b), which can be fitted by MAPbI2.5 with ordered iodine vacancies as shown in Fig. 5j. Compared with the pristine MAPbI3 in Fig. 5i, part of iodine ions on the (001) plane indicated by the black circles in Fig. 5j are lost to form the superstructure MAPbI2.5. The corresponding simulated ED pattern of MAPbI2.5 in Fig. 5f matches well with the experimental result in Fig. 5b. Furthermore, from another viewing direction of [100] of perovskite, the ED pattern corresponding to superstructure MAPbI2.5 is also observed (see details in Supplementary Fig. 11). In fact, iodine ions are the most likely mobile ions due to the lower diffuse barrier in MAPbI311,42,43, similar to the oxygen ions in the oxide perovskites44, and the migration of iodine ions in MAPbI3 indeed has been reported from the previous studies10,14,45. Thus, we infer that the superstructure is likely caused by the formation of ordered I vacancies. Moreover, forming such an intermediate superstructure MAPbI2.5 tends to minimize the repulsive interaction between iodine ions as well as the elastic coherence strain energy, similar to the staging phases in the lithium ion battery materials46,47.

With further increased electron beam dose, the structure continues to lose MA and additional I ions to form MAyPbI2.5–z (0 ≤ y ≤ 1 and 0 ≤ z ≤ 0.5) (Fig. 5k), which is accompanied by the disappearance of the superstructure. Finally, the perovskite structure framework collapses to form the layered PbI2 structure (Fig. 5l). Note that the ED pattern in Fig. 5c can be regarded as a combination of MAyPbI2.5–z along [001] and PbI2 along [$$\bar 4$$41] directions, consistent with the overlap of the corresponding simulated ED patterns in Fig. 5g (see details in Supplementary Fig. 12). It is also interesting to note that the previous study has proposed the reaction pathway from PbI2 to MAPbI3 during synthesis48, and the layered structure of MA1+xPbI3+y was proposed as the intermediate phase, similar to our observation of MAyPbI2.5–z.

To better understand the decomposition process from MAPbI3 to PbI2, we also investigate the intensity changes of diffraction spots during the decomposition as shown in Fig. 5m. Under the electron beam illumination, the intensities of the MAPbI3 (220), (110), and (200) spots decrease, indicating that the perovskite structure becomes defective. At 73 s, the (100) superstructure reflection starts to appear likely due to the formation of MAPbI2.5. With the growing of MAPbI2.5, its (100) reflection intensity gradually increases. After a peak intensity of MAPbI2.5 (100) at 328 s, MAPbI2.5 further loses I and MA to form MAyPbI2.5–z (0 ≤ y ≤ 1 and 0 ≤ z ≤ 0.5), leading to the reduction of superstructure reflection intensity. Meanwhile, the perovskite structure collapses and decomposes into a layered PbI2 structure. After 933 s, the superstructure MAPbI2.5 has completely disappeared. And after 1524 s, the reflections of MAyPbI2.5–z (0 ≤ y ≤ 1 and 0 ≤ z ≤ 0.5) phase disappear and the intensity of the (0$$\bar 1$$4) spot of PbI2 is stabilized, indicating that the structure has completely decomposed into PbI2.

Finally, we investigate the effect of the dose rate on the degradation, and we determine the critical electron dose limit to obtain the intrinsic perovskite structure. Note that the perovskite framework remains until the MAyPbI2.5–z (0 ≤ y ≤ 1 and 0 ≤ z ≤ 0.5) collapses. Therefore, it is reasonable to expect that before the formation of MAPbI2.5 superstructure, the perovskite structure with a low density of iodine vacancies is robust and close to the ideal pristine structure. We find that typically the formation of MAPbI2.5 starts at 303 s under a dose rate of 0.5 e Å−2 s−1 (total dose: 151 e Å−2) and 73 s under a dose rate of 1 e Å−2 s−1 (total dose: 73 e Å−2), while at 2 and 4 e Å−2 s−1, the decomposition has already begun at the first SAED images (~30 s of exposure) (see details in Supplementary Fig. 13). Note that the decomposition not only depends on the total dose, but is also sensitive to dose rate, with higher rates resulting in accelerated decomposition under a similar total dose.

## Discussion

The MAPbI3 is extremely sensitive to electron irradiation and thus it is easy to decompose into PbI2 accompanied by subtle changes in ED, i.e., some of the diffraction spots are missing or weaken, or additional spots appear. Particularly, if we can only observe {2h, 2k, 0} diffraction spots along the [001] direction while the {2h+1, 2k+1, 0} reflections [e.g., (110)] are absent, it is highly likely that the perovskite structure has already decomposed into PbI2. However, many previous studies ignored such subtle changes and incorrectly identified the decomposed PbI2 to be MAPbI3 based on the ED or FFT analysis. Consequently, extra attention must be paid during analyzing the electron microscopy data for these materials.

More importantly, we find that the decomposition of MAPbI3 is likely initiated with the loss of I ions to subsequently form superstructure MAPbI2.5, followed by the loss of MA and additional I ions to form MAyPbI2.5–z (0 ≤ y ≤ 1 and 0 ≤ z ≤ 0.5) accompanied by the disappearance of a superstructure, and finally decomposes into PbI2. Typically, under electron beam illumination with a dose rate of 0.5 e Å−2 s−1, the MAPbI3 starts to change into MAPbI2.5 with ordered I vacancies after 303 s, leaving sufficient time to acquire the ED for an intrinsic perovskite structure, while a higher dose rate results in faster decomposition under a similar total dose. With a dose rate higher than 2 e Å−2 s−1, decomposition has already begun at the first SAED images. These identified TEM conditions can be used to guide future electron microscopy characterization of these materials.

Moreover, considering the similarities between electron beam illumination and optical illumination, as well as the heating and electric field, during which the ions gain energy to overcome the migration barriers to trigger the phase transitions, we believe that the electro-driven ion migration and heating will likely lead to a similar structural degradation process, which may be responsible for the poor long-term device stability. In this sense, the observed decomposition of perovskite under the electron beam illumination can help us to understand the structure instability issue for these materials under device operation, which is currently under further investigation.

## Methods

### MAPbI3 synthesis

The single-crystal perovskite MAPbI3 was synthesized as follows. Two FTO/TiO2 substrates were clamped together using the fixed-size card slots. In total, 2 mL of precursor solution of PbI2/MAI (molar ratio 1:1) was prepared in γ-butyrolactone. After 12 h of continuous stirring, a homogeneous MAPbI3 solution at a concentration of 1.3 mol/L was obtained, and then filtered using a polytetrafluoroethylene (PTFE) filter with 0.22-µm pore size. The clamped FTO/TiO2 substrates were then vertically dipped into a 10-mL beaker containing MAPbI3 precursor solution kept on a hot plate at 120 oC in nitrogen atmosphere, and the feeding precursor solution was added every 12 h. After a certain period of 27 days, the substrates with a self-grown film were taken out, and then dried at 120 oC for 10 min in nitrogen glovebox.

### Data acquisition and analysis

Powder XRD patterns were obtained on D8 Advance diffractometer using Cu Kα radiation (40 kV and 40 mA) with a scanning rate of 4° min−1 for wide-angle test increment. The morphologies of the sample were examined by SEM (ZEISS Gemini SEM 300), and the SEM-EDX was carried out at 10 kV with 1 min for acquiring the spectra, using a current of ~ 1 nA. The beam was kept blank for several minutes after each spectrum was acquired because the sample is sensitive to the electron beam.

Single-crystal X-ray diffraction was carried out at Sector 7-ID-C, Advanced Photon Source, Argonne National Laboratory, using 10 -keV X-ray with a 300 × 300 μm2 beam size as defined by a slit. A Huber six-circle diffractometer coupled with a PILATUS 100-K area detector was employed for measurements of the single-crystalline samples. The detector was placed downstream from the samples such that ~ 8° coverage in the 2θ-angle and ~ 3° in the χ-angle was obtained. Rocking (ω-angle) scans around each reflection were recorded and for the data presented, the resultant 3D data volume was reduced by integrating the rocking angles.

High-resolution TEM, HAADF STEM images were acquired at an aberration-corrected FEI (Titan Cubed Themis G2) operated at 80 kV equipped with an X-FEG gun and Bruker Super-X EDX detectors. STEM images were acquired with a beam current of 6–10 pA, a convergence semi-angle of 25 mrad, and a collection semi-angle snap in the range of 53–260 mrad. The STEM-EDX mapping was obtained with a beam current of 1 nA and counts ranging from 2k cps to 8k cps for ~ 5 min. The SAED patterns for Fig. 5, Supplementary Fig. 11, and Fig. 13 were acquired at 300 kV.

Under the STEM mode for Figs. 2 and 3, the dose rate is estimated to be 220–360 e Å−2 s−1 which is calculated by dividing the screen current (6 ~ 10 pA) by the area of the raster49. To get a satisfactory STEM image, we need to carefully align the zone axis and adjust the astigmatism and the focus before image acquisition, which is time-consuming and costs about 5–10 min. To record the SAED patterns of Fig. 1d, the sample is always observed at low magnifications (10000–30000) and the dose rate is estimated to be 100–300 e Å−2 s−1. It usually takes 30–100 s to record a SAED image after the sample comes into sight. Under these conditions, judging from the data we acquired, the samples were decomposed due to large electron doses. To obtain the ideal pristine structure, we make the dose rate as low as possible. For Fig. 5, Supplementary Figs. 11 and 13, the dose rate ranges from 0.5 to 4 e Å−2 s−1.

The simulations of the ED pattern were performed by using Crystalmaker software, and the ball-and-stick models were formed using Vesta software. The FFT and inverse FFT patterns were obtained using DigitalMicrograph (Gatan) software. HAADF images in Fig. 2, Fig. 3, and Supplementary Fig. 5 were averaged from multiple regions with a homemade MATLAB code to reduce noise. The plots were drawn using Origin 2018.

### STEM simulation

STEM simulation was carried out by Kirkland with COMPUTEM software50. Specifically, the technique involves dividing the sample into a number of thin slices normal to the incident electron beam and calculating the contribution to the cross-section at each slice. First, we used Vesta software to construct a 15 × 15 × 5 supercell containing 12,352 atoms based on the PbI2 CIF file39 and a 5 × 5 × 5 supercell containing 41,387 atoms from the MAPbI3 CIF file51. Then a homemade MATLAB code was used to rotate the structure to the desired zone axes. The STEM simulated parameters were set with the beam energy of 80 keV (accelerating voltage), object aperture of 25.1 mrad (convergence semi-angle), and a STEM ADF detector of 53–260 mrad (collection semi-angle). Besides, the transmission function size was 2048 × 2048 pixels and STEM probe size was 512 × 512 pixels. The other parameters were set as default. To achieve realistic computing time for these configurations, a limited area was chosen to run the STEM simulations. The thickness of the sample along different zone axes was estimated to range from 4.5 to 6.5 nm.

### Ab initio simulation

All ab initio simulations were performed using Vienna Ab Initio Simulation Package52,53,54 projected augmented wave (PAW)55,56 potential. For calculations of PbI2, strongly constrained and appropriately normed (SCAN)57 exchange-correlation functional was adopted. Plane-wave cutoff of 650 eV was used. For structural optimization, 6 × 6 × 6 k-mesh was used. For molecular dynamics simulations, 1 × 1 × 1 k-mesh and 81-atom supercell were used. Isothermal–isobaric ensemble (NPT)58 simulations at 300 K were run 5000 steps to obtain averaged lattice constants and subsequently, canonical ensemble (NVT) simulations at the same temperature were run 10,000 steps using Nosé–Hoover thermostat59,60, of which the final 8000 steps were used to average the atomic positions. The time step was chosen to be 2 fs for both simulations and the energy drift was less than 1 meV (ps)−1 per atom for NVT simulations.