a SEM image of the red InGaN nanorod array coated with an epitaxial passivation layer. b The schematic of e-AlN formation by Ar plasma-assisted ALD. MQW (R), SLs, and p-EBL represent red InGaN/GaN multiple quantum wells, InGaN/GaN superlattices, and p-type AlGaN electron blocking layer, respectively. c Atomic arrangement at the passivated interface between e-AlN and non-polar GaN or InGaN. d Corresponding equilibrium bandgap diagram of (c). e Timing sequence of the Ar plasma-assisted ALD for e-AlN (0.5–10 nm) + n-Al₂O₃ or n-SiO₂ (40–49 nm). TEMAH, TMA, and DIPAS represent tetrakis(ethylmethylamino) hafnium (hafnium precursor), trimethylaluminium (aluminum precursor), and di-isopropylamino silane (silicon precursor), respectively. A scale bar, 2 µm (a).

The e-AlN is fabricated by the Ar plasma-assisted ALD (Fig. 1e). Although ALD is often used as a passivation method for complex 3D structures due to its superior conformality, ALD AlN layers are typically amorphous or polycrystalline. Here, we realise the epitaxial growth of ultrathin single crystalline AlN on the sidewalls of nano- and μ-LEDs. Our approach features introduction of the Ar plasma ignited between ALD growth cycles^15,16, as compared to that of conventional ALD (Supplementary Fig. 1a). The in-between Ar plasma generates local heating that provides thermal energy to arriving precursors for the atomic rearrangements on (In)GaN LED sidewall surfaces to form epitaxial AlN (Fig. 1b). In detail, the process is designed to ignite the Ar plasma after every 4-ALD-cycles (Fig. 1e and Supplementary Fig. 1b) to minimize the surface roughness of the passivation layer while maintaining an epitaxial relationship. The details of ALD timing sequences are described in “Methods”. It should be mentioned that the nano- and μ-LED m-plane sidewalls vertical to the substrate surface are much more beneficial for epitaxial passivation compared to horizontal c-face epilayers. This is because there was little Ar plasma-induced physical damage on the vertical sidewalls with nearly parallel to the direction of ion bombardment.

To prove the concept of epitaxial passivation, we first grew e-AlN on c-plane p-type GaN using the suggested method. Both high-resolution transmission electron microscopy(HR-TEM) and atomic high-angle annular dark-field scanning TEM (HAADF-STEM) images clearly reveals hetero-epitaxy between c-plane GaN and AlN (Fig. 2a–f). However, inspection of the figures shows that with increasing thickness in the range of 0.8–5.6 nm, the e-AlN exhibits uneven contrast (Fig. 2a–c) or poor lattice ordering (Fig. 2d–f). This suggests that the crystalline quality of the thicker e-AlN may be lower due to lattice mismatch. Next, we coated a InGaN-based nanorod LED array with various passivation layers to form core/shell structure (Supplementary Fig. 2). The n-HfO₂/AlN/Al₂O₃/SiO₂ represents normal or conventional amorphous HfO_x/AlN/Al₂O₃/SiO₂ ALD without Ar plasma heating (Supplementary Fig. 1a). The vertical and horizontal cross-section of the e-AlN coated nanorod (Fig. 1a) is characterized by HR-TEM (Fig. 2g–m). Inspection of Fig. 2g–i with coating of e-AlN (10 nm)/n-AlN (40 nm) reveals there exist hetero-epitaxy between AlN and (In)GaN m-plane sidewalls. We note that the e-AlN layer appeared to extend to more than 10 nm, probably because of the nature of the n-AlN that was easily crystallized. When combined with other amorphous layers, such as SiO₂ and Al₂O₃, e.g., AlN/SiO₂ and AlN/Al₂O₃, 10-nm e-AlN was clearly visible (Fig. 2j–m). However, Fig. 2g–i again shows uneven contrast, suggesting existence of defects for the 10-nm-thick e-AlN passivation layer due to lattice mismatch between (In)GaN and AlN. Importantly, for ultrathin (~2-nm-thick) e-AlN, we can observe conformal and clean AlN/(In)GaN hetero-epitaxial relationship on m-plane sidewalls (Fig. 2j–m). We claim that the suggested ALD process enables such ultrathin e-AlN on the sidewall, importance of which is ability to make the thickness close the critical thickness (t_{cr, AlN}), below which the hetero-interface is in principle coherent without any dislocation.

Fig. 2: Proof-of-concept of e-AlN under different passivation/substrate conditions.

a–f HR-TEM (a–c) and atomic HAADF-STEM (d–f) images of e-AlN on c-plane GaN films with variable thicknesses of 0.8 (a, d), 2,3 (b, e), and 5.6 nm (c, f). g–i TEM images of the red InGaN/GaN nanorod LEDs coated with e-AlN (10 nm)/n-AlN (40 nm). g, h Vertical cross-sections near MQWs and n-GaN, respectively. i Horizontal cross-section near MQWs. Insets of (h, i), Corresponding SAD pattern. j–m TEM images of the red LED nanorods coated with e-AlN (~2 nm)/n-Al₂O₃ (48 nm). j, k Vertical cross-sections near MQWs (j) and n-GaN with FFT pattern (k). l, m Horizontal cross-sections near MQWs at two different locations. HAADF-STEM, HR-TEM, SAD, and FFT represent high-angle annular dark-field scanning transmission electron microscopy, high resolution TEM, selected area diffraction, and fast Fourier transform, respectively. Scale bars, 1 nm (a–f), 5 nm (g–m), and 5 nm⁻¹ (Insets of e–g).

The critical shell thicknesses for wurtzite or zinc blende GaN/Al_xGa_1-xN and GaN/In_yGa_1-yN core/shell hetero-structure can be calculated^17,18. Dislocation-free strain accommodation in the core-shell nano-structure that can effectively increase t_{cr, AlN} is available only when the GaN core diameter is below ~10 nm. Therefore, in our case with the diameter of GaN nanorod core larger than 600 nm, t_{cr, AlN} should be close to the bulk value of conventional film hetero-epitaxy, which is estimated to be ~1 nm. In case of AlN/In_0.35Ga_0.65N hetero-epitaxy in the red QW, t_{cr, AlN} should be lower than 1 nm due to larger lattice mismatch. To verify the concept, we further in-depth analysed the structure of 2-nm-thick e-AlN on the InGaN/GaN MQW sidewall using atomic HAADF-STEM (Fig. 3a–e), which is able to distinguish heavy In or Ga atoms from light Al atoms by bright and dark contrast, respectively. As estimated, we can identify that AlN/GaN interface is fully coherent while, in the AlN/InGaN interface, there is a defective (low atomic ordering) region (Fig. 3a) or dislocation (Fig. 3b). Inspection of Fig. 3a, b shows the thickness of e-AlN is 5–7 atomic layers, corresponding to 1.5–2.1 nm. Interestingly, we observe dislocation in InGaN QW from the surface probably induced by e-AlN (Fig. 3c). We note that, due to the cylindrical shape of the nanorod, vertical cross-sectional observation often exhibits curved-edge artefact image unless the sampling thickness is thin enough (Fig. 3d and Supplementary Fig. 3a). Nevertheless, the atom-layer-thick line profile analysis, where only a single atomic layer is included with a width of 0.25 nm (Supplementary Fig. 3b), clearly reveals gradually decreasing oscillation amplitude abruptly decreases in the e-AlN region (Supplementary Fig. 3c). We can identify that e-AlN is 5-Al-atom-layer thick, corresponding to the thickness of AlN is ~1.5 nm, consistent with that in Fig. 3a, b. The atom-layer-thick line profile analysis might be useful to get rid of the artefact in the TEM analyses of ultrathin film on curved surface. On the other hand, horizontal cross-sectional TEM enables the curved artefact-free observation on ultrathin e-AlN (Fig. 3e), though it is somewhat difficult to identify point of interest. Moreover, energy dispersive spectroscopy (EDS) analyses compositionally verified the 2-nm-thick e-AlN on InGaN/GaN MQW sidewalls (Fig. 3f). It is reported that the required thickness of dielectric layer to determine its Q_f to be fixed is 5 ALD cycle (~0.5 nm)¹⁹. We claim that 1- or 2-nm-thick e-AlN, with high crystallinity except presence of slight dislocation or defects near InGaN QW, is ideal chemical and physical passivation.

Fig. 3: In-depth atomic structure analyses of ~2-nm-thick e-AlN at MQW region.

a–d Atomic HAADF-STEM images of the e-AlN. a–d Vertical cross-sections at different locations, showing AlN-(In)GaN hetero-epitaxy. a, b Conformal e-AlN with fully coherent epitaxy at e-AlN/GaN interface. At e-AlN/InGaN interface, defective (low atomic ordering) region (a) and dislocation (b) observed in e-AlN, marked by arrows. c Dislocation induced in InGaN marked by an arrow. d Conformal e-AlN on MQWs with thicker TEM sampling. (inset) corresponding FFT image. e, Horizontal cross-sections near MQWs. f (left) HAADF-STEM image. Next panels are corresponding EDS maps. Scale bars, 2 nm (a, b), 5 nm (c, d), 5 nm⁻¹ (Inset of d), 1 nm (e), and 7 nm (f).

The epitaxial passivation effect itself was optically compared with other surface passivation including n-HfO₂ and n-AlN for the InGaN/GaN nanorod LED arrays (diameter/pitch = 0.6 μm/1.8 μm). Fluorescence microscopy (Fig. 4a–d) revealed that the e-AlN passivation showed superior photoluminescence (PL) intensity and uniformity compared to that by n-AlN or n-HfO₂. The PL intensity of the nanorod array obtained by epitaxial passivation was 2.5 times larger than that of n-HfO₂ (Fig. 4e). We note that epitaxial passivation showed superior PL intensities over a wide thickness range (1–30 nm) as shown in Supplementary Fig. 4. We suggest that the first few nanometres of AlN determined the PL enhancement, while subsequent n-AlN or n-Al₂O₃ layers acted only as insulating layers for EL devices. It can be seen that the PL enhancement factor for epitaxial passivation was approximately 2.5 (Supplementary Fig. 4). We also compared the PL decay characteristics (τ_PL) for the same samples with time-resolved PL (TRPL, Fig. 4f). As shown in the figure, we obtained a fourfold increase in the τ_PL^20,21 passivated by e-AlN compared to that passivated by n-HfO₂ as summarized in Supplementary Table 1; this indicated enhancement of the internal quantum efficiency (“Methods”). Next, we prepared the red nanorod-LED arrays with more versatile passivation layers and compared their relative electroluminescence (EL) efficiencies (η_EL) in Fig. 4g. Fabrication flow of the nano-LEDs is described in Supplementary Fig. 5a with device structures (Supplementary Figs. 5b, c), exhibiting microscopically non-uniform colour/intensity variation (Fig. 4h). It is evident that most e-AlN/n-Al₂O₃ based nano-LEDs yields higher efficiencies than that of the control LEDs (n-HfO₂/n-Al₂O₃ and n-AlN/n-Al₂O₃) for full range of the injection current, where n- Al₂O₃ (40–49 nm) is used for insulating the LEDs. This confirms beneficial influence of epitaxial passivation, consistent with optical data (Fig. 4d–f). As discussed in the structural analyses, the t_AlN has striking effect on η_EL. In the range of t_AlN of 0.5–2 nm, the η_EL of the nano-LED with e-AlN is drastically increased, ~2.5 times larger than that with n-HfO₂ for the whole current injection range. The maximum η_EL is achieved at t_AlN of 1 nm with slightly lower values at t_AlN of 0.5 nm and 2 nm. However, at t_AlN of 10 nm, η_EL decreases again significantly. The efficiency decrease with increasing t_AlN is consistent with discussion in the structural analyses in Figs. 2 and 3, associated with high defect generation in the e-AlN much beyond t_{cr, AlN}. On the contrary, there is very few thickness effect of e-AlN on the optical intensity (Fig. 4d–f), which is under investigation.

Fig. 4: PL and EL of red nano-LEDs with various ALD passivation layers.

Fluorescence microscopy images of the as-dry/wet etched state (a), coated with n-HfO₂ (10 nm)/n-Al₂O₃ (40 nm) (b), with n-AlN (10 nm)/n-Al₂O₃ (40 nm) (c), and with e-AlN (10 nm)/n-Al₂O₃ (40 nm) (d). e, f Time-resolved PL measurements. PL spectra (e) and decay time curves (f) using. g EL efficiency comparison of the nano-rod LEDs for various passivation layers. h Typical optical microscope image of the nano-rod LEDs. Scale bars, 10 µm (a–d, h).

By choosing the optimal e-AlN (e-AlN (1 nm)/n-Al₂O₃ (49 nm)) determined from Fig. 4g, we fabricated the flipped InGaN/GaN μ-LEDs (diameter/pitch = 1.5 μm/4.5 μm) for high efficiency, with detailed process flow in Supplementary Figs. 6a–i. We note that the fabricated red μ-LEDs (Supplementary Fig. 6j) seem to exhibit macroscopically uniform EL emission even at low current density of 0.24 A per cm² and high brightness with increasing current (Fig. 5a and Supplementary Fig. 6k). However, we observe that there is still microscopic EL non-uniformity at low current densities in the range of 0.8–3.2 A per cm² (Fig. 5b). Considering that even the e-AlN passivated nano-LEDs show quite uniform PL intensity (Fig. 4d), we attribute the EL non-uniformity of the μ-LEDs to non-uniform current injection. To address this issue, we are investigating the non-uniform ohmic ITO contact with p-type GaN. It should be mentioned that the non-uniform current injection is much stronger in nano-LEDs. Inspection of Fig. 4h reveals local pink colour, mixture of red and blue peaked at ~460 nm (Fig. 5d), suggesting strong localized current injection. This strong colour variation/localized current injection in nano-LEDs could be associated with randomly distributed V-pit defects of comparable size to nano-LEDs, which are commonly observed in red epitaxial layers. For the larger diameter μ-LEDs, microscopic colour variation is significantly reduced probably due to decrease in V-pit induced local current injection effect (Fig. 5b).

Fig. 5: EL of red InGaN/GaN μ-LEDs with e-AlN (~1 nm)/n-Al₂O₃ (49 nm).

a Low-magnification optical microscope (OM) images in the current range of 0.06–5 mA. b High-magnification OM images in the current range of 0.2–0.8 mA. c EQE and peak wavelength (λ_peak) curves vs. current density (J) before and after roughening with tetramethylammonium hydroxide (TMAH). d Corresponding emission spectra of the cyan curve in left panel. e EQE and λ_peak curves vs. J for four different wafers after the TMAH roughening. f Converted EQE vs λ_peak relations at current densities of maximum EQE and of ~40 A per cm² for six different samples (18 points), including samples in (e). Scale bars, 200 µm (a) and 20 µm (b).

The flipped μLEDs, without applying encapsulating polymer, are designed to also maximize the extraction efficiency (EE) by surface roughening and adequate placement of Al reflection layer (Supplementary Fig. 6i). It was found that surface roughening achieved by the TMAH wet treatment increases the EE by ~1.5 times (Fig. 5c). Despite the EQE variation depending on the wafer position, the peak absolute EQE of red μ-LEDs is 5.8% at the peak wavelength (λ_peak) of 645 nm and the current density of ~1 A per cm² (Fig. 5c), measured by integrating sphere (“Methods” and Supplementary Fig. 7). The EQE is ~3.1% even at relatively high current density (J) of 40 A per cm². As spectral broadening and blueshift is inevitable like other green/red InGaN LEDs with blue side peaks (Fig. 5d), absorption type filter would be required for display application. To investigate wafer to wafer variation, we have further measured absolute EQE and λ_peak for four different wafers with slightly different emission characteristics (Fig. 5e). For #1-#3 wafers, the peak EQEs were 6.1–7.0% with λ_peak of 646–649 nm at J of 0.2–0.8 A per cm². The median EQE of 6.5% with λ_peak of 649 nm is shown in the cyan curve in Fig. 5e and Supplementary Fig. 6l. To our knowledge, this red EQE value has been significantly improved for μ-LEDs of such a small diameter of 1.5 μm. On the other hand, for #4 wafer, relatively low peak EQE of 4.5% is observed with λ_peak of 653 nm at J of 2.0 A per cm². Considering its high initial λ_peak of 670 nm at J of 0.2 A per cm², which is 11–16 nm larger than those of #1-#3 at the same J, it is estimated that #4 has higher indium concentration (x) in In_xGa_1-xN QW. This explains the observed low EQE due to increased defects in the high x QW. At J of 40 A per cm², absolute EQEs exhibit smaller variation in the range of 2.8–3.3% (Fig. 5e). It should be mentioned that the EQE convergence into the narrower range at high current density is attributed to the fact that Auger recombination dominates the droop characteristics of the μ-LEDs^22,23. To improve EQE uniformity and droop characteristics, optimization of the InGaN LED epitaxial film is also needed. We get statistical data on EQEs of 18 points (six different wafers, including four wafers in Fig. 5e and 3 different points per wafer) using the spectra-radiometer. (Fig. 5f) The spectro-radiometer measures optical power for the given solid angle to calculate relative EQE, which we convert to absolute EQE by multiplying the conversion factor, as described in “Methods” and Supplementary Fig. 7. The peak converted EQEs at J of 0.3–2 A per cm² were 4.5–7.0% with λ_peak of 646–649 nm, and the converted EQEs at J of 40 A per cm² were 2.2–3.4% with λ_peak of 619–629 nm, exhibiting the similar tendency as in Fig. 5e. As explained, there is decreasing tendency of the peak EQE with increasing peak wavelength, marked by a dotted line in Fig. 5f.

Near complete transfer for flexible micro-LEDs

In this section, we demonstrate flexible inorganic UHD displays with their fabrication scheme shown in Fig. 6a–c; it can transfer red μ-LEDs, called a near-complete device. As described in Supplementary Fig. 8 and “Methods”, the fabricated near-complete device is flipped and bonded to a PI-coated glass substrate with an adhesive polymer (Fig. 6a), followed by removal of the growth substrate (Fig. 6b). The transfer scheme has the distinct ability to facilitate the peel-off-free removal of the Si growth substrate, which is inherently controlled by the interface strength. When we divide the total grinding force (F) during mechanical removal of the Si substrate into the normal grinding force F_n and tangential grinding force F_t, we see that peel-off of the PI film occurs during mechanical grinding only if F_t is larger than the adhesion strength or force (G) of the PI (flexible substrate)/glass (carrier substrate) interface, F_t > G. In the suggested fabrication scheme, the above condition can be satisfied spontaneously; therefore, peel-off rarely occurs during grinding, as described below. F_t monotonically increases with increasing wafer rotation speed, which increases the abrasive wheel feed rate and decreases the wheel rotation speed²⁴. On the other hand, systematic steered molecular dynamics simulations were used to show that G also increased with sliding speed²⁵. It was also reported that the absolute value of G was maintained at a high level with a moderate choice of PI with higher stiffness and low thermal expansion coefficient and by controlling the processing parameters such as UV radiation or PI curing time²⁶. These studies explain why the devised scheme is highly robust. In other words, F_t and G increase together with increasing sliding speed while holding F_t < G, which makes the process invulnerable to grinding parameters, including the wheel or wafer rotation speed. This peel-off-free removal with a large process window will be verified in the following sections. Low-alignment-accuracy post processing, including the common electrode to the n-GaN (N-comm) and pad opening, finalizes the flipped, near-complete device, which becomes flexible via excimer laser lift off (LLO, Fig. 6c), as described in the “Methods”. With the detailed fabrication flow (Supplementary Fig. 8), we verified flexible red μ-LEDs. It can be seen that the flexible μ-LEDs maintain highly efficient, bright, and macroscopically uniform EL emission both in the middle of the release step (Fig. 6d, e) and in the fully released flexible form (Fig. 6f, g).

Fig. 6: Near-complete device transfer for flexible e-AlN passivated red μLEDs.

a–c Schematic fabrication flow. a Flipping of the red μLEDs to enable adhesive bonding on PI/glass. b Epi-wafer removal by controlling the interfacial forces and subsequent n-GaN removal. c Postprocessing including pad electrode formation, and laser lift-off. d, e Operation of flexible red μ-LEDs in the middle of the release step; photograph (d) and microscope image (e). f Photograph image of released μ-LEDs from a carrier glass substrate corresponding to (c). g Photograph images of flexible red μ-LEDs operating under fully flexible forms. h Conceptual schematic for an electronic contact lens with a flexible display. A scale bar, 200 µm (e).

For the reader’s reference, we have also demonstrated flexible 5000 ppi displays using the planar blue InGaN μ-LEDs pixelated by ion implantation^3,27,28. We show three changeable still images (zebra, checker, and gradation) at the selection of V_DD1, V_DD2, and V_DD3 with the corresponding pad electrodes (Supplementary Fig. 9). This is achieved by designing the unique TFT-free structure on the planar blue μ-LEDs with the selective Via1 opening pattern, as described in the “Methods” and Supplementary Fig. 10a–m. The TFT-free assembly can be transferred onto a PI/glass substrate, post-processed, and released into a flexible form. Supplementary Fig. 10w shows the robust interfaces formed after the transfer. The released flexible displays, with a total thickness of less than 20 μm, exhibited stable electroluminescence (EL) spectra (Supplementary Fig. 9c), and defect-free, changeable still images at variable bending forms (bottom panels of Supplementary Fig. 9a). This 5000 ppi pixel density was verified by optical microscopy images at the fixed location with various magnifications (Supplementary Fig. 9d). We expect that the mesa-etched red InGaN μ-LEDs, once flattened by chemical mechanical polishing, can also be integrated into the UHD display using the same fabrication process (Supplementary Fig. 10n–v). We believe that with the development of monolithic integration of both high-mobility channel-based TFTs and RGB μ-LEDs, the described structure will evolve into a monolithic flexible UHD full-colour μ-LED display. We hope that this will pave the way for electronic contact lens²⁹, an eventual form factor for near-eye display (Fig. 6h). Lastly, we describe our preliminary progress on monolithic TFT integration and lens assembly in the supplementary information (Supplementary Figs. 11–17).

We have demonstrated 1-nm-thick e-AlN passivation of 1.5 μm diameter μ-LEDs with the EQE of 6.5% at 649 nm by developing Ar plasma heating-based ALD. The e-AlN passivation almost perfectly passivates the tiny red μ-LEDs with conformal thickness of 1 or 2 nm, as verified by HAADF-STEM analyses. We then have demonstrated flexible red InGaN μ-LEDs by developing the near complete device transfer process. By breaking the red spectral efficiency and form factor bottlenecks for inorganic μ-LEDs, we believe our study can provide a paradigm for highly efficient, flexible UHD RGB displays that can expand the range of AR applications, including electronic contact lenses.