Fig. 1. Artificial intelligence applied in various fields of full-color display: (a) epitaxial structure design, (b) defect detection and repair, (c) synthesis of perovskite, and (d) dimming. Figures reproduced with permission from (a) Ref. 31, under a Creative Commons Attribution (CC-BY) license; Ref. 32, under a CC-BY license; (b) Ref. 33, under a CC-BY license; Ref. 34, under a CC-BY license; (c) Ref. 35, under a CC-BY license; Ref. 36, under a CC-BY license; (d) Ref. 37, under a CC-BY license.

Fig. 2. (a) Schematic diagram of the micro-LED backlight module: (i) schematic diagram of the LED with DBR structure; (ii) highly reflective surface substrate; (iii) etching structure of the receiver. (b) The workflow of the environmental control agent includes the implementation of micro-LED backlight module prototypes and the use of a CMOS sensor to capture images for optimizing simulation parameters. (c) The process of establishing optical simulation models involves using R-soft and LightTools software to simulate micro-LED backlight modules with various DBR structure configurations. Figures reproduced with permission from Ref. 32, under a CC-BY license. (d) The principles of Gaussian and Lambertian scattering modes, which are used to simulate the reflective characteristics of light in micro-LED backlight modules. (e) The properties of the bidirectional scattering distribution function (BSDF), which records the intensity and angular distribution of reflection scattering and refraction scattering produced by rays at various incident angles on the film stack. (f) During the DRL optimization process, the changes in iteration uniformity are demonstrated through comparative images, corresponding to reward function 1, reward function 2, and reward function 3, respectively. (g) Under high-resolution conditions, the best uniformity results achieved by DRL are presented, corresponding to reward function 1, reward function 2, and reward function 3, respectively. Figures reproduced with permission from Ref. 32, under a CC-BY license.

Fig. 3. (a) Flow chart of the entire prediction framework based on machine learning (ML), the whole process contains four main parts: data collection, preprocessing, model selection and analysis, and evaluation metrics. (b) A sketch map of the input features. The composition, doping concentration, and structural parameters in MQW, EBL, n-GaN, and p-GaN of the InGaN blue LEDs are selected as features of each sample. In the abbreviation, W, B, and E before the hyphen represent quantum well, quantum barrier, and EBL layer, respectively. Figures reproduced with permission from Ref. 45, under a CC-BY license. (c) The impact of various parameters on GaN-LEDs’ IQE: (i) Heatmap of the performance deterioration degree, with the numbers in squares representing the exact values of performance deterioration. (ii) SHAP summary plot for the importance ranking. The x-axis position of the dot is its SHAP value, where a more positive SHAP value indicates that this feature value has a larger positive impact on the model output (IQE in this figure) and vice versa. (iii) 4D scatter plot of IQE with respect to W-T, B-T, and E-T. (iv) 2D contour map of IQE predictions when B-T = 10 nm, showing that W-T has a greater impact on IQE than E-T. Figures reproduced with permission from Ref. 45, under a CC-BY license.

Fig. 4. (a) Schematic diagram of the mini-LED backlight module and single packaged LED. (b) Workflow of the DDQN algorithm. (c) Overall design of the optical simulation and algorithm. (d) Schematic diagram of modeling in LightTools. (e) The mini-LED backlight module’s simulation model built in LightTools software. (f) The illumination uniformity results of the mini-LED backlight module after optimization with the DDQN algorithm. (g) A comparison of illumination uniformity under different optimization parameters. Figures reproduced with permission from Ref. 46, under a CC-BY license.

Fig. 5. (a) Screening framework for two-dimensional silver/bismuth (2D AgBi) iodide perovskites. (b) Problem-specific descriptors for predicting the synthesis feasibility of two-dimensional silver/bismuth iodide perovskites, including the count of nitrogen atoms, steric effect index, interatomic distance, eccentricity, and count of rotatable bonds. (c) ROC curve and confusion matrix for the SVC model with an AUC value of 0.85. (d) SHAP values indicating the contribution of different descriptors to the prediction results in the ML model. (e) SHAP scatter plot revealing the relationship between descriptors and synthesis feasibility, showing the impact of molecular eccentricity and rotatable tail bonds. (f) SHAP bar chart illustrating the influence of chemical groups on the synthesis feasibility of two-dimensional silver/bismuth iodide perovskites, with colors indicating enhancement or weakening of synthesis feasibility. (g) Illustration of the commercial availability of materials, with 123 being commercially available and 221 being commercially unavailable. In addition, 344 two-dimensional perovskites and 8062 non-two-dimensional perovskites were predicted. (h) Displays of physical images of several two-dimensional silver/bismuth iodide perovskites, including (C6H11NH3)4AgBiI8, (FC6H4CH2NH3)4AgBiI8, (FC6H4C2H4NH3)4AgBiI8·H2O, (CIC6H4CH2NH3)4AgBiI8, (BrC6H4CH2NH3)4AgBiI8, (C4H2C5H6NH3)4AgBiI8·H2O, and (NH3C6H4CH2NH3)2AgBiI8. Figures reproduced with permission from Ref. 36, under a CC-BY license.

Fig. 6. (a) The SEResNet model, enhanced by the feature mask (FM) method, is used to optimize the performance prediction of perovskite solar cells (PSCs), including PCE, VOC, JSC, and FF. (b) The schema of the PI algorithm. Starting with the material properties and structural characteristics of perovskite solar cells, the process involves data preprocessing, machine learning model training, using the PI algorithm to assess feature importance, and finally experimentally validating the model’s predictive performance. (c) The key characteristics affecting the PCE of perovskite solar cells identified by the PI algorithm, in which Pb, I, and MA are most important, providing the direction for material optimization. (d) PCE statistical diagram of the actual experimental device. (e) The trend comparison chart between experimental data and predicted data showing six different PSC samples. Figures reproduced with permission from Ref. 59, under a CC-BY license.

Fig. 7. (a) Dehydrogenation mechanism: (i) dehydrogenation in the contact anneal process before source (S)/drain(D) deposition and (ii) vertical structure after the S/D process. (b) Graph of ion experimental values based on the distance between the contact hole and the driving TR. (c) Transistor TEG image and parameter extraction for AI model training: (i) transistor at the center of the image, with C1, C2, C3, and C4 marking the contact hole locations; (ii) boundary extraction from the experimental image using the Canny edge-detection algorithm; (iii) extraction of main inflection-point parameters from the image, with key inflection points denoted by blue dots. (d) Mura simulation results using the image-quality prediction system: (i) real mura images; (ii) simulated images from the AI model. Figures reproduced with permission from Ref. 33, under a CC-BY license.

Fig. 8. Schematic diagram of the YOLO-ADPAM framework, including the backbone, neck, and prediction components, highlighting the critical roles of the PAM, ASPP, and DSCM modules in feature extraction and fusion. Figures reproduced with permission from Ref. 70, under a CC-BY license.

Fig. 9. (a) Schematic diagram of local dimming. (b) Simplified dual-panel display system for local dimming. (c) Illustration of dimming processes: (i) original image; (ii) global backlight brightness distribution; (iii) pixel-compensated image after global dimming; (iv) one-dimensional backlight brightness distribution; (v) pixel-compensated image after one-dimensional dimming; (vi) local dimming backlight brightness distribution; (vii) pixel-compensated image after local dimming.

Fig. 10. (a) Conventional local dimming process for LCDs. (b) Overall architecture of the proposed local dimming system. The proposed LDNN can be implemented (i) on the TV-set side or (ii) on the LCD module side. (c) The LDNN’s hourglass-shaped architecture features upper blue skip connections that concatenate convolutional layer data to upsampling layers, and lower black skip connections that add data. Blue indicates strided convolution layers for downsampling, whereas green marks strided transposed convolution layers for upsampling. Non-marked convolution layers have a stride of 1, with layer numbers indicating spatial resolutions. (d) Demonstrating the proposed LDNN’s effectiveness, this figure shows test results on three images with the following stages: (i) input image (IIN): the original image; (ii) local dimming (LD): backlight dimming effect; (iii) compensated image (ICP): post-compensation enhancement; (iv) output image (IOUT): final image post-LDNN processing. The LDNN’s ability to preserve image quality is evident through the clear and detailed IOUT images. Figures reproduced with permission from Ref. 37, under a CC-BY license.