Introduction

Optoelectronic devices, which generate electrical signals using light or vice versa, have garnered significant attention for their diverse applications in fields such as optical communication1,2, medical analysis3,4,5, astronomy6, environmental monitoring7,8, and energy systems9. Among the key performance parameters of optoelectronic devices, transmittance (\(\:T\)) plays a crucial role because it directly affects the overall efficiency by regulating the intensity of the incident light. Consequently, enhancing the transmittance and improving the device performance have become major focal points in optoelectronic research and industry. In this regard, harnessing nanostructures has emerged as a promising approach for increasing the transmittance of optoelectronic devices. Nanostructures offer advantages in enhancing the transmittance through a high surface-to-volume ratio, providing a simple and reproducible means of improving the transmittance while suppressing the reflection and scattering of incident light10,11,12,13. These structures have also been explored in research involving solar cells and OLEDs, where their ability to control light propagation plays a crucial role in improving device efficiency14,15,16. Thus, various studies optimising the nanostructure and improving the transmittance of optoelectronic devices have been conventionally reported based on numerical simulations, such as rigorous coupled-wave analysis (RCWA), the finite element method (FEM), and the finite-difference time-domain (FDTD)17,18,19,20.

More recently, there has been growing interest in research that incorporates data-based artificial intelligence (AI) technology, including machine learning and deep learning, for optimising nanostructures21,22,23,24,25,26,27. This innovative approach involves converting the results obtained from numerical or theoretical analyses into big data, allowing the calculation of the relationship between the input and output. The use of AI technology enables the accurate prediction of optical properties, even non-linear behaviours, and facilitates the prediction of continuous outputs, including the optical properties of untrained nanostructures. As a result, the prediction of optical properties becomes more efficient, significantly reducing time consumption. Moreover, the bidirectional result-prediction capability of AI allows for an inverse design24,25,26,27,28,29,30,31,32,33, enabling the identification of the necessary structure based on the desired outcome. This enhancement in efficiency and usability further supports optimal structure design.

In this context, AI-based technology has the potential to improve the transmittance using nanostructures. However, certain considerations are still crucial for its practical utilisation: (1) Highly efficient development of an optimised learning model for the desired optical response: Despite the development of numerous AI-based models, the prediction performance of the model can vary significantly depending on the form of the learning data. Unless an optimised learning model for a specific dataset is developed, it may require a myriad of data to define the relationship between the input and output data, implying cost and time consumption. Thus, given the diversity of AI models and the settings of subvariables, it is essential to develop and utilise a model that can accurately and efficiently predict results, even when provided with limited data. (2) Consideration of fabrication tolerance: Although the design of an optimal nanostructure is possible, the 100% precise fabrication of such structures using practical processes is challenging. Thus, it is necessary to predict the performance range of nanostructures that exhibit the desired optical properties and provide information on fabrication tolerance. This will enable researchers to account for uncertainties in the fabrication process and design nanostructures that are more practical for implementation.

To address these issues, various studies have recently been conducted. Deep learning-based approaches can optimise the parameters for the desired optical response and predict targeted optical properties with high accuracy; however, they still inevitably require a significant initial cost of large amounts of training data28,29,30,31,32,33,34,35, indicating significant time and energy expenses and making the system complex. Furthermore, to the best of our knowledge, most previous studies have focused solely on optimising parameters and improving prediction accuracy; thus, few studies have focused on integrating a method to provide sufficient information considering the targeted performance margin with fabrication tolerance, which is very important for implementing the designed nanostructure for practical applications. Herein, we propose a simple method for improving the performance of optoelectronic devices by developing an efficient and accurate AI-based nanostructure prediction model. The developed model not only accurately predicts the optimal nanostructure for high transmittance in a specific range, such as the visible-light range, with high efficiency but also provides fabrication tolerance. An accurate and efficient machine learning model was developed based on support vector regression (SVR), which optimised the associated hyperparameters. Because the developed model is based on SVR, which allows a margin in the learning process, we could achieve a higher accuracy with only hundreds of data points that can offer significantly higher efficiency and accuracy than that calculated by an artificial neural network (ANN). Importantly, the developed model exhibits superior predictive performance even for untrained dimensions of the nanostructure, allowing for contour mapping of continuous results. Thus, it has the advantage of detecting fabrication tolerance by providing a structural range with similar transmittance. To validate the practical effectiveness of the proposed method, we demonstrated the optimal nanograting structure in the visible-light range by utilising the developed method and applying these dimensions to optoelectrical devices such as solar cells and OLED. Consequently, the optimal nanostructure not only showed the highest transmittance experimentally, but the nanograting structure associated devices also showed a significant performance enhancement corresponding to the transmittance improvement.

Results and discussion

SVR-based AI model development

Figure 1 shows an overall schematic of the proposed method. The proposed model involves the continuous training of nanostructure-transmittance data obtained from FDTD simulations (Fig. 1-i, ii) and visualising it on a contour plot (Fig. 1-iii, iv), thus enabling the efficient prediction and optimisation of nanostructure-transmittance, including the performance range and fabrication tolerance. To develop the prediction model, we attempted to exploit an SVR-based machine learning method36,37 that can efficiently predict the transmittance accurately with minimal data. Specifically, SVR is an effective regression model based on support vector machines (SVM) that utilises kernel functions to find hyperplanes in a higher-dimensional space, thereby facilitating the analysis of input data and accurate prediction of nonlinear input-output relationships, including optical properties. In particular, when predicting the input-output relationship, the implementation of margin settings allows for efficient regression with high accuracy, even with relatively limited data, demonstrating high effectiveness. Therefore, we utilised limited data (n = 216) for training but achieved high accuracy in the transmittance with respect to different dimensions of the nanograting structure. Furthermore, through the concept of margins in training, the proposed model can accurately predict continuous data involving untrained conditions and provide information regarding the range of similar results in the trained nanostructure-transmittance, ensuring fabrication tolerance. Using the trained results, we developed a contour mapping visualisation method that defines the same transmittance range with a closed line for the nanostructure-transmittance results, which allows us to predict the range of nanostructures suitable for the desired optical properties.

Fig. 1
figure 1

Schematic of the proposed concept. 3D schematic of the proposed optical device, utilising a nanograting structure optimised by the proposed SVR model. The developed SVR-based model shows high efficiency and accuracy because it is well optimised, and also provides fabrication tolerance (Range in iv) Contour map.

Full size image

Figure 2 shows the development of the SVR-based machine learning model in detail. To develop the proposed model, we first gathered various transmittance data corresponding to different nanograting structures at a specific wavelength (\(\:\lambda\:\)=850 nm) for the training set. More specifically, we set the dimensional variables of the nanograting as pitch (P), duty (D), and height (H) (Fig. 2a) and obtained more than two-hundred transmittance data points using the FDTD simulation (Fig. 2b). Specifically, transmittance simulation values were obtained at points of P= 200–400 nm (interval = 40 nm), D= 0.4–0.8 nm (interval = 0.08 nm), H= 200–400 nm (interval = 40 nm), and the total number of data was 216 (n = 216). Moreover, an unpolarized light source was used for the simulation. The details of the FDTD simulation are presented in Supplementary Information, Figure S1.

Fig. 2
figure 2

Development of the proposed SVR model. (a) Schematic of the proposed nanograting film (Inset: dimensional definition of nanograting structure). (b) FDTD simulation model. (c) Training sets for the SVR model calculated by FDTD simulation (\(\:\lambda\:\) = 850 nm). (d) Performance comparison between SVR, KNN, and GR. (e) Performance comparison of the developed SVR model with linear kernel function and polynomial kernel function. (f) Performance comparison of the developed SVR with a conventional ANN. (g) Continuous transmittance prediction according to duty from the developed model. (h) Predicted continuous transmittance map across various ranges of nanograting dimensions (P = 200 nm, \(\:\lambda\:\:\)= 850 nm). White dots represent learning data, the position and color of each dot indicating structural variables and transmittance information, respectively. (i) Verification results of the proposed SVR model with randomly selected nanograting dimensions at 450, 650, and 850 nm wavelength.

Full size image

The obtained FDTD data regarding the transmittance (T) is plotted in 4-dimensional space, including the structural variable axes (P, D, and H) (Fig. 2c). However, it is difficult to confirm a clear regularity or linear relationship between the structural variables (P, D, and H) and transmittance (T). To define the specific relationship, we first adapted a few training models, such as k-nearest neighbour (KNN), SVR, and Gaussian regression (GR), which are representative models for calculating nonlinear relationships (Fig. 2d). We confirmed a radial basis function (RBF) kernel-based SVR model shows the highest accuracy (R-square (R2) = 0.995 and mean squared error (MSE) = 0.004), compared to that of KNN (R2 = 0.971 and MSE = 0.022) and GR (R2 = 0.833 and MSE = 0.130). When using the Linear kernel or Polynomial kernel instead of the RBF kernel, we observed lower accuracy (R2 = 0.183 and MSE = 0.636 @ Linear kernel, R2 = 0.126 and MSE = 0.680 @ Polynomial kernel) in comparison to the RBF kernel (Fig. 2e). We understand that the highest accuracy is because the RBF kernel function should not only have intrinsic advantages in analysing the nonlinear relationship, but also other hyperparameters that are well optimised. Conventionally, the RBF kernel-based SVR model has three hyperparameters, epsilon (\(\:\epsilon\:\)), cost (C), and gamma (\(\:\gamma\:\)) which represents the width of the margin, noise (or penalty, equivalently), and influence of individual training data, respectively. To optimise the hyperparameters, we conducted grid search, then, we confirmed the highest accuracy (R2 = 0.995 and MSE = 0.004) with \(\:\epsilon\:\) = 0.1, C = 20, and \(\:\gamma\:\) = 0.15 (Supplementary Information Figure S2). We also conducted a study on the change in the accuracy of the developed model with respect to the amount of training data (Supplementary Information Figure S3). Considering the amount of data, we set the level of the uniform interval for the variables of the nanograting structure (P, D, and H) to 2, 3, 5, and 6 for the set level, generating 8, 27, 125 and 216 training data, respectively. Consequently, R2 and MSE increased and decreased, respectively, as the number of data points increased, implying an improvement in the model accuracy. The highest model accuracy was achieved with 216 data points.

The efficiency and accuracy of the developed model were verified. To verify the efficiency of the developed model with high accuracy, we further compared the performance of the developed model, such as R2, MSE, and training time (\(\:\tau\:\)), with that of the training model developed by the artificial neural network (ANN) (Fig. 2f). For the comparison, we developed a conventional ANN-based training model using 216-data and confirmed the performances of the developed SVR model (R2SVR = 0.995, MSESVR = 0.004, and \(\:{\tau\:}_{SVR}\:\)= 8.380) show notably higher accuracy and time (and cost, equivalently) efficiency than that of ANN (R2ANN = 0.909, MSEANN = 0.071, and \(\:{\tau\:}_{ANN}\) = 840.973), with a limited dataset (Supplementary Information Figure S4). The details of the developed ANN model are provided in Supporting Information Table S1.

Developed model verification

Subsequently, we verified the developed SVR model. For verification, we first visualised a continuous T map with respect to various nanograting dimensions, such as D, H, and P. As previously mentioned in Fig. 1, the developed SVR model can predict T of the untrained nanograting dimension based on its algorithm; thus, we can continuously plot T by adapting the developed model for D (Fig. 2g). Then, by combining various T-D plots with respect to H, we obtain the continuous T map with respect to various nanograting dimensions (P = 200 nm, D = 0.4–0.8, H = 200–400 nm) (Fig. 2h). We also confirmed the possibility of using the developed model to predict T with respect to nanograting dimensions for different \(\:\lambda\:\) values. For that, we repetitively obtained the 216-data about T with respect to various D, H, and P for different \(\:\lambda\:\) (450 and 650 nm), then adapted the developed model and visualised the continuous T map with same protocol. The model verification is conducted by randomly selecting six nanostructures from the mapped T at each \(\:\lambda\:\) = 450, 650, and 850 nm, and fitted the predicted T (Tpred) of the randomly selected nanostructures with the T (Tsimul) obtained by additional FDTD simulation (Fig. 2i). As a result, the Tpred and Tsimul show high correspondences (R2 = 0.949 and MSE = 0.062), demonstrating the high accuracy of the developed SVR model to predict T about wide range of nanograting structure and incident light \(\:\lambda\:\). The predicted T maps about \(\:\lambda\:\) = 450 and 650 nm and the randomly selected data points are shown in Supplementary Information Figure S5.

Prediction of the fabrication tolerance

Based on the fact that the developed model accurately predicts the T of various nanograting structures, we further investigated the nanograting structure that provides the maximum T in the visible-light range for high-performance optical devices. To optimise the nanograting structure in visible-light range, we first theoretically predicted the T maps of various nanograting structures (P = 200 nm, D = 0.4–0.8 and H = 200–400 nm) at \(\:\lambda\:\) = 400–1000 nm (interval = 50 nm) (Fig. 3a). Then, by averaging the whole T value with respect to the various dimensions, we could theoretically achieve the specific nanograting structure to maximise the T in the visible-light range (T\(\:\:\ge\:\:\)91.8 @ P = 200 nm, D = 0.38–0.56 and H = 230–292 nm). The T mapping data for P, D, H, and \(\:\lambda\:\) are shown in Supplementary Information Figure S6. Importantly, we can utilise the predicted T map to obtain the optimised T within the fabrication tolerance. Because the developed model continuously predicts T with respect to different dimensional parameters, contouring can be conducted using the same T value on the predicted T map. Moreover, the contoured line forms a closed loop, implying that the nanograting dimensions in a specific contour line have higher T than the structures of that line (T\(\:\:\ge\:\:\)91.8 with in the solid line in Fig. 3b). Thus, the contouring map can be utilised for fabrication tolerance. It is worthy to be noted that a more detailed example of the performance of transmittance affected by fabrication tolerance is shown in Supplementary Information Figure S7.

Fig. 3
figure 3

Nanograting optimisation in visible-light range and experimental verification. (a) Predicted transmittance maps of various dimensions in visible light range (λ = 400 ~ 1000 nm) and. ( b)The result of averaging transmittance across wavelengths from 400 to 1000 nm. The black solid lines represent contours of equal transmittance. (c) Fabrication process for the proposed polymer nanograting film. (d) Optical image of fabricated polymer nanograting film. (e) SEM images of three different nanograting structures; (i) P = 200 nm, D = 0.5, H = 250 nm, (ii) P = 200 nm, D = 0.5, H = 350 nm, (iii) P = 400 nm, D = 0.625, H = 350 nm. The PUA nanograting is coated with platinum (Pt) for ion beam milling. (f) The measured transmittance spectra of fabricated nanograting films. (g) Predicted transmittance spectra of nanograting structures calculated by the developed SVR model. (h) Comparison of average values in the visible-light range of measured and predicted transmittances.

Full size image

Experimental verification

To experimentally verify the optimal T, we fabricated various optically transparent nanograting structures using polyurethane acrylate (PUA) and polyethylene terephthalate (PET). The fabrication process was developed by combining a conventional semiconductor process with nano-imprint lithography (Fig. 3c and Methods). Because the developed process is based on techniques that are well established at the industrial level, it enables the demonstration of the targeted nanograting structure over a wide area (8-inch in, Fig. 3d). Three different nanograting structures with theoretically different T values (Supplementary Information Figure S8) were fabricated (P = 200 nm, D = 0.5, H = 250 nm, P = 200 nm, D = 0.5, H = 350 nm, P = 400 nm, D = 0.625, H = 350 nm; Fig. 3e) using different Si templates for comparison (Supplementary Information Figure S9).

The experimental T (Texp) of the fabricated nanograting structure was measured using UV-VIS spectroscopy (Fig. 3f). It is worthy to be noted that a polarizer was not applied in the set-up to consider the unpolarized light. All the specimens exhibited fluctuating transmittance, but the trends were different. Specifically, the nanograting structure of P = 200 nm, D = 0.5, H = 250 nm shows higher T (> 90%) in the visible-light range (400–1000 nm) than that of others; the nanograting of P = 200 nm, D = 0.5, H = 350 nm shows slightly lower T at \(\:\lambda\:\:\)< 500 nm and the nanograting of P = 400 nm, D = 0.625, H = 350 nm shows sudden T decrease at \(\:\lambda\:\:\)< 700 nm, respectively. We also compared the measured values with the T calculated using the developed model (Tpred, Fig. 3g) and confirmed that the trends of Texp and Tpred corresponded well, indicating the high accuracy of the developed model. For a quantitative comparison, we averaged Texp and Tpred of each nanograting specimen over the visible-light range and confirmed the high correspondence between Texp and Tpred (Fig. 3h). It should be noted that the experimentally demonstrated PUA nanograting (ii. P = 200 nm, D = 0.5, H = 350 nm) has some dimensional imperfection (D = 0.55 & 0.45), but this imperfection is negligible in theoretical and experimental transmittance-wavelength results (Supplementary Information Figure S10).

Applications for optoelectronic devices

Using an optimised nanograting structure in the visible-light range, we fabricated solution-processed organic light-emitting diodes (SP-OLEDs) based on flat and nanograting substrates utilising a green emitter (TCTA: TPBi: Ir(mppy)3). Remarkably, the incorporation of the nanograting substrate into the SP-OLEDs significantly enhanced the device performance. Figure 4(a) shows a schematic diagram illustrating the multilayer structure of the nanograting enhanced SP-OLEDs, which can be described as follows: ITO as the bottom electrode/PEDOT: PSS as the HIL/TCTA: TPBi: Ir(mppy)3 as the EML/TPBi as the ETL/LiF as the EIL/Al as the top electrode. An energy band diagram of this structure is shown in Fig. 4(b). When a voltage was applied to the SP-OLED, electrons were injected from the top electrode (Al) into the ETL composed of the TPBI film. Holes were simultaneously injected from the bottom electrode (ITO) into the HIL, which consisted of a PEDOT: PSS film layer. The injected electrons and holes migrated towards the EML, which was formed by the TCTA: TPBi: Ir(mppy)3 composite. Within the EML, organic molecules are excited by the recombination of electrons and holes, resulting in the formation of excitons. The excitons formed in the EML undergo radiative decay and emit photons of specific wavelengths corresponding to the energy levels of the EML material. This emission generated the desired light output from the OLED38,39,40. After exciton formation and emission, any remaining electrons and holes not involved in the radiative recombination process continued their transport through the ETL and HIL, ensuring a continuous flow of charge carriers. The ETL, TPBI, aids in efficient electron transport and facilitates the injection of electrons from the top electrode. The EIL and LiF further promote electron injection by lowering the electron injection barrier38,39,40,41,42. Overall, the combination of specific materials and the enhanced performance provided by OLEDs with nanograting substrates contributes to improved device efficiency and enhanced charge carrier balance, ultimately resulting in enhanced SP-OLED performance. The SP-OLEDs with nanograting substrates demonstrated a notable increase in luminance and current density, particularly in the high-voltage region, as shown in Fig. 4(c), which shows the current density-luminance-voltage (J-V-L) characteristics. The Von at a constant luminance of 1 cd/m2 was measured at 4.30 V for the SP-OLEDs with the nanograting substrate indicating that it operates at a voltage lower than 4.50 V of the reference device. The enhanced SP-OLEDs with nanograting exhibited a luminance of 17,038 cd/m2, indicating a significant improvement of 16% compared with the luminance of the reference device, which was 10,729 cd/m2. Both the reference and nanograting devices demonstrated green electroluminescence (EL), with a maximum spectral wavelength (λmax) of 510 nm, as shown in Fig. 4(d). The color coordinates, determined by the Commission Internationale de l’Eclairage (CIE) system, were (0.29, 0.62) for the reference device and (0.29, 0.62) for the nanograting devices, as shown in Supplementary Information Figure S11. Remarkably, despite the presence of the nanograting substrate, the emission wavelength remained unaffected, as no deviation was observed in the peak wavelength of the reference device.

Fig. 4
figure 4

(a) Schematics of flexible OLED structure. (b) Energy level diagram of OLED device. (c) Current Density-Voltage-Luminance, (d) Electroluminescence (EL) spectra and CIE 1931 chromaticity diagram, (e) External quantum efficiency-Current density, and (f) Current efficiency-Current density-Power efficiency.

Full size image

The SP-OLEDs with nanograting also yielded improvements in the EQE (Fig. 4(e)), current efficiency (CE), and power efficiency (PE) (Fig. 4(f)), corresponding to the improvement in the J-V-L characteristics. The SP-OLEDs with nanograting exhibited a maximum EQE of 6.86%, indicating a significant improvement of 17% compared with that of the reference device, which was 5.85%. The SP-OLEDs with nanograting achieved the highest CE of 26.53 cd/A and PE of 11.35 lm/ W. The SP-OLED parameters are summarised in Table 1. To characterise the potential use of the fabricated nanograting substrate as an electrode in organic solar cells (OSCs), we prepared OSCs with the following device structure: PET with nanograting/ITO/ZnO nanoparticles/photoactive layer/PEDOT: PSS/Ag (Fig. 5(a)). The photoactive layer is composed of a bulk heterojunction blend consisting of PTB7 as the electron donor and PC71BM as the electron acceptor. For comparison, reference OSCs with flat ITO films (flat) on PET and PET anodes were fabricated under similar conditions. The highest occupied molecular orbital (HOMOs) and lowest unoccupied molecular orbital (LUMOs) energy levels of the materials used in this study are shown in Fig. 5(b). The current density-voltage (J-V) characteristics of the ITO/PET film and ITO/nanograting on the PET film-based OSCs under AM 1.5G illumination are shown in Fig. 5(c). The average values for the ten samples are listed in Table 2. In general, the nanograting-based OSCs exhibited a slightly higher performance than the devices. For example, as can be seen from Fig. 5(c), an ITO film-based OSC (flat) with a PTB7:PC71BM photoactive layer exhibited a short-circuit current density (Jsc) of approximately 16.45 mA/cm2, an open-circuit voltage (Voc) of approximately 0.73 V, and a fill factor (FF) of approximately 64.61%, resulting in a PCE of 7.74 ± 0.12%. Conversely, the nanograting-based OSCs had a reasonably high PCE, of 8.57 ± 0.05% (Jsc ≈ 17.15 mA/ cm2, Voc ≈ 0.74 V, FF ≈ 67.69%). The higher Jsc values of the OSCs fabricated using ITO/nanograting on PET films could be mainly attributed to the higher incident-photon-to-current conversion efficiency (IPCE) of the OSCs because the optical transmittance of the electrodes was greater for wavelengths between 450 and 750 nm (Fig. 5d).

Table 1 Performance of flexible OLEDs.
Full size table
Fig. 5
figure 5

(a) Schematics of flexible organic solar cells (OSCs). (b) Energy level diagram of OSCs. (c) J-V curves of OSCs with different substrates. (d) IPCE spectra of OSCs without or with Nanograting.

Full size image
Table 2 Performance of flexible organic solar cells (averaged 10 samples).
Full size table

Conclusion

In this study, we proposed an efficient and accurate AI-based nanostructure design optimisation method to enhance the performance of optoelectronic devices. The SVR-based machine learning model that includes the concept of margin enables the development of a high-performance prediction model that shows an MSE of 0.004 and an R2 of 0.995 with only 216 learning data points.

The developed model demonstrates superior accuracy compared to the ANN model, which is the most basic deep learning model, and indicates remarkable efficiency by requiring 100 times less learning time. Moreover, the developed model predicts the continuous transmittance within the range of specific structural variables to determine the optimal nanostructure that satisfies the target optical responses. The nanostructure-transmittance contour map, which visualises the prediction results of the developed model, can define the range of nanostructures that satisfy the target transmittance and provide information on the fabrication tolerances. Utilising this method, a PUA-PET nanograting film designed for visible-light ranges exhibited a high similarity of trends between the experimental values and model predictions, implying its potential for practical device applications.

The results of device performance when using the fabricated nanograting film, a 17% improvement in the EQE of SP-OLEDs, and a 10.7% improvement in the PCE of OSC, verified the effectiveness of the application at the device level. Therefore, the proposed method is expected to not only efficiently and accurately design the optimal nanostructure for the target optical response but also improve the reliability of the actual fabrication step by providing fabrication tolerance. Furthermore, the results of its application to actual devices open up possibilities for a wide range of optoelectronic devices, beyond LEDs and SC.

Materials and methods

Nanograting PUA film fabrication

The fabrication process began with a silicon (Si) nanograting template fabricated using conventional KrF lithography. A typical nano-imprint lithography process was employed to fabricate the nanograting PUA substrate. First, UV-curable PUA (HC11M-J5, Minuta Technology Co., Ltd., South Korea) resin was spin-coated on the Si nanograting template for 30 s at 1000 rpm and covered with 50-µm-thick PET film. The PUA was cured using a mask aligner (MDA-400LJ, MIDAS System Co., Ltd., South Korea) with an I-line (365 nm wavelength) at 50 mJ/cm2. Finally, the PUA film with a replicated nanograting structure was fabricated by peeling the polymer substrate from the template. The same process was repeated on three different structures of Si nanograting templates (Supplementary Information, Figure S9) to fabricate nanograting PUA films with three different structures (Fig. 3e).

Electrode fabrication

The ITO/PET electrodes were vacuum-sputtered in our laboratory. The ITO/PET films were formed by depositing a layer of ITO on a 50 μm-thick PET substrate (SKC Co. Ltd.) via radio-frequency (RF) superimposed DC magnetron sputtering. An ITO target comprising 10 wt% tin was used for sputtering. The ITO sputtering was performed using an Ar gas flow rate of 30 sccm, an O2 gas flow rate of 0.3 sccm, and a working pressure of approximately 1.5 × 10 − 1 torr. The distance between the substrate and plasma source was approximately 100 mm. The DC power was controlled by keeping the current constant (approximately 0.5 A) and an RF power of approximately 50 W was imposed simultaneously. The growth rate of the film was approximately 0.316 nm s-1, and an ITO film with a final thickness of approximately 100 nm was deposited.

Organic solar cell fabrication

The electrodes used as anodes in the OSCs had the following structure: PET/ITO/ZnO NPs/PTB7:PC71BM/PEDOT: PSS/Ag. A 5 nm thick layer of ZnO NPs was spin-coated onto the electrodes. The coated buffer layer was cured at a temperature of 80 °C for 5 min. A 100 nm thick active layer of PTB7:PC71BM was spin-coated using a mixture of PTB7:PC71BM (8 mg: 12 mg) dissolved in 1 ml of chlorobenzene and mixed with 3 vol% 1, 8-diiodooctane (Sigma Aldrich). PTB7 and PC71BM were purchased from 1-material and Solemme BV, respectively. PEDOT: PSS (Clevios™ PVP AI 4083, Heraeus Epurio, Germany) diluted with isopropyl alcohol (IPA) at a PEDOT: PEDOT-PSS: IPA ratio of 1:10 was deposited onto the photoactive layer. A 100 nm thick Ag electrode with a defined area of 0.36 cm2 was deposited via evaporation on the active layer.

Organic light-emitting diode fabrication

PET/ITO substrates with a sheet resistance of 15 Ω per square were washed in the order of acetone, distilled water, and isopropanol. After ozone treatment for 10 min, PEDOT: PSS was first spin-coated onto the ITO glass substrates with a spin rate of 4000 rpm and annealed at 100 °C for 10 min. The device structure was ITO/PEDOT: PSS(40 nm)/EML(80 nm)/TPBi(30 nm)/LiF(1 nm)/Al(100 nm). The EML, which was made of a mixed host and dopant Ir(mppy)3(93:7 wt% relatively) in CB, was deposited at 1000 rpm onto the PEDOT: PSS layer and then annealed at 80 °C for 5 min. TPBi, LiF, and Al (40, 1, and 100 nm thick, respectively) were deposited by thermal evaporation at a base pressure of 1 × 10− 6 Pa.

Characterisation

The photocurrent density-voltage (J-V) properties of the OSCs were measured using a solar simulator (Newport 450 W 3 A solar system, Oriel Instruments, USA) with AM 1.5 illumination as the light source. The calibrated light intensity was set to 1 Sun, using a standard silicon cell. The effective area of the device was defined as the projected area transmitted by the mask, which was equal to the diameter of the photoanode multiplied by its length (0.36 cm). Incident photon-to-current conversion efficiency (IPCE) was measured using the incident photon-to-current conversion efficiency measurement system (QuantX 300, Oriel Instruments, USA) with an Oriel Cornerstone™ 130 1/8 m monochromator operated in AC mode. The device performance and electroluminescence (EL) spectra of the green OLEDs were measured using an M3000 parameter test system.