In recent years, full-color display technology has emerged as a cornerstone of the information era, driving advancements in visual experiences across a wide range of cutting-edge applications. From ultra-high-definition displays in scientific imaging and medical diagnostics to immersive visualization in virtual reality (VR) and augmented reality (AR), full-color displays provide an unparalleled window into the digital world.1^–3 Renowned for their exceptional color fidelity and dynamic visual performance, these technologies have become essential tools in pushing the frontiers of human–machine interaction and data representation. At the heart of full-color display technology is the precise reproduction of the natural spectrum of colors, accomplished through the controlled integration of red, green, and blue (RGB) primary colors, supported by advanced color management algorithms.4^–6

Initially, liquid crystal displays (LCDs) dominated the display market with their ability to produce high-resolution images at a relatively low cost. However, the advent of organic light-emitting diodes (OLEDs) marked a significant leap forward, offering superior contrast ratios and faster response time, making them ideal for high-end applications. More recently, the emergence of micro-LEDs and quantum LEDs (QLEDs) has further pushed the boundaries of display technology. For example, Yu et al.7 reported the development of AlGaN-based deep ultraviolet micro-LEDs emitting at 275 nm, which significantly enhanced the performance of high-fidelity displays. In addition, Sun et al.8 demonstrated enhanced ultraviolet luminescence of AlGaN wavy quantum well structures grown on large misoriented sapphire substrates, further advancing the material science behind full-color displays. However, as demands for higher resolution, broader color gamuts, superior contrast, and energy-efficient operation intensify, traditional display technologies are increasingly strained to meet these expectations.9^–11 In this context, the evolution of full-color display systems not only is a matter of innovation in material science but also requires the adoption of intelligent algorithms to optimize color accuracy, energy management, and overall system efficiency.12

Since the groundbreaking fabrication of GaN-based micro-LEDs by Jin et al. in 2000, the field of full-color display technology has undergone revolutionary advancements.13 In 2011, Hong et al.14 developed color-tunable LEDs spanning red to blue by utilizing InGaN/GaN multi-quantum wells (MQWs) on GaN nanorods. In 2015, Han et al.15 integrated quantum dots with micro-LEDs via aerosol jet printing and enhanced light utilization efficiency with distributed Bragg reflectors (DBRs). By 2017, Lin et al.16 successfully reduced optical crosstalk in quantum dot-based full-color micro-LED displays using photolithographic resist molds, thereby improving color purity and contrast. In 2019, Kuo et al. employed atomic layer deposition and nonradiative resonant energy transfer techniques to improve the efficiency of hybrid quantum dot nanoring micro-LEDs.17 Subsequently, in 2021, Lin et al. demonstrated printable quantum dot photopolymers as color conversion layers, achieving precise quantum dot deposition and full-color displays through inkjet printing. In 2022, Hwangbo et al.18 introduced a wafer-scale monolithic integration approach, directly growing two-dimensional semiconductor materials (${\mathrm{MoS}}_2$) on GaN epitaxial wafers to achieve fully integrated micro-LED structures. Most recently, in 2023, Qi et al.19 presented monolithic full-color active matrix micro-LED displays by heterogeneously integrating InGaN and AlGaInP materials. Their design incorporated flip-chip bonding with CMOS backplanes to form a dual-layer thin-film structure, successfully realizing full-color imaging.19 In 2024, Memon et al.20 introduced a multifunctional three-terminal diode that integrates a traditional GaN-based p–n diode with a third terminal, enabling it to function as both a light emitter and a detector by combining these capabilities into a single structure. These advancements collectively highlight the rapid progress and transformative potential of the full-color micro-LED technology, paving the way for future breakthroughs in high-performance displays. The progress of micro-LEDs and their applications, as detailed in Table 1, underscore this transformative potential.

Artificial intelligence (AI), recognized as one of the most transformative technological revolutions of the 21st century, is profoundly reshaping numerous fields, including autonomous vehicles, medical diagnostics, financial risk management, and creative industries. In the photonics domain, AI has emerged as a key driving force behind innovations in full-color display technologies, leveraging its exceptional data processing and pattern recognition capabilities. By analyzing vast color datasets, optimizing display panel calibration, and enabling personalized and context-aware color adjustments, AI is significantly enhancing user experiences.22^,23

The groundbreaking role of AI in scientific research was underscored in 2024 when its applications were honored with a Nobel Prize. This milestone not only recognized AI’s pivotal contributions to scientific advancement but also highlighted its profound potential in driving innovation. Today, AI’s integration into full-color display technologies has expanded beyond early applications in color management to encompass intelligent control systems.24^–26 Through deep learning and machine vision, AI has accelerated the design of high-performance epitaxial structures, revolutionized defect detection and automated repair processes, and ensured stability and reliability in display manufacturing. Furthermore, in quantum dot (QD) display technologies, AI enables precise synthesis and alignment of QDs, achieving unprecedented levels of color precision and fidelity.27 In addition, AI’s role in driving circuit design, advanced packaging techniques, and dynamic dimming systems has significantly enhanced the overall performance of display products, propelling full-color display technologies toward greater intelligence and efficiency.28^,29

Although significant progress has been made in integrating AI for material growth, structure design, and engineering, there are still few studies discussing the use and effect of AI during the fabrication processes of a specific application such as full-color displays.30 As shown in Fig. 1, to systematically elucidate the potential of AI in this domain, this review comprehensively explores AI-driven innovations in perovskite synthesis, defect management, and application-specific optimization. Key processes, such as AI-guided synthesis and precision defect repair, are examined to highlight their impact on enhancing material quality and device performance. Finally, we identify the opportunities and challenges for AI integration, emphasizing its pivotal role in driving the future of intelligent, sustainable, and high-performance full-color displays.

In epitaxial structure design, blue and deep ultraviolet (DUV) LEDs, though primarily focused on single-wavelength emissions, play a crucial role in full-color display technologies. Particularly, the combination of these LEDs with quantum dots is an essential approach for achieving full-color displays. For example, blue LEDs combined with quantum dots can efficiently convert light to enable the broad color spectrum required for full-color displays. In addition, DUV LEDs provide a more efficient excitation source for quantum dots, further enhancing display performance. Therefore, the optimization of blue and DUV LEDs in epitaxial structure design is not only aimed at improving single-wavelength emission efficiency but also aimed at enabling breakthroughs in full-color display technologies.38

In recent years, the development of epitaxial structures for full-color display technology has progressed rapidly, playing a pivotal role in enhancing display performance, stability, and efficiency.39 AI-driven epitaxial structure design, leveraging the deep integration of machine learning models, enables precise prediction and optimization of key parameters in full-color display LEDs (e.g., group-III nitride GaN structures).40 These parameters include quantum well width, barrier layer material and thickness, and the composition of the electron-blocking layer, significantly accelerating the discovery process for high-efficiency LED structures. Compared with traditional experiments, AI-driven optimization of epitaxial structures for full-color displays offers substantial advantages.41^,42 Based on a large number of experimental data and simulation results, AI technology could establish optimized predictive models to significantly improve the efficiency and accuracy of epitaxial design, thereby reducing the time and cost required by traditional experiments.16^,30

Gallium nitride (GaN)-based LEDs have garnered significant attention due to their widespread applications in lighting, displays, and other fields.43 Optimizing their epitaxial design is crucial for enhancing device performance and reducing production costs. In 2016, Rouet et al.31 proposed an active machine learning strategy to optimize GaN-based LEDs, which significantly improved electroluminescence efficiency at high current densities. By simulating GaN-based LED structures, the approach achieved a uniform carrier concentration distribution in the active region, effectively delaying the onset of efficiency droop. The research team precisely controlled carrier distribution by adjusting the indium (In) content in the quantum wells and the design of the quantum barriers. This design not only influenced carrier transport properties but also addressed the regulation of defect states within the material. By varying the In content in the quantum wells, researchers were able to adjust the potential depth of the wells, thereby impacting carrier distribution. Wider quantum wells facilitated carrier dispersion under high current densities, reducing localized carrier accumulation in the active region and delaying efficiency droop. In addition, optimizing the indium content in the quantum barriers modulated the carrier tunneling probability, further promoting a uniform carrier distribution across the active region. This active learning strategy rapidly developed a model for predicting device performance using Poisson–Schrödinger simulations while generating structures with higher simulation efficiency. Compared with conventional trial-and-error simulations, this method significantly reduced time requirements. Furthermore, the efficiency of the machine learning-optimized structures under high current densities was nearly 40% higher than that of reference LED structures.

In 2021, Huang et al.32 proposed an innovative method to optimize micro-LED backlight module design using deep reinforcement learning (DRL) combined with a micro-macro hybrid environment control agent. This approach addressed the complex computational environment of micro- and macro-scale structures in micro-LED backlight module design, which traditional methods often struggled to handle accurately. The research team developed a virtual-reality workflow incorporating an environment control agent to ensure high consistency between design environment parameters and experimental results. They precisely designed the thickness and material composition of the DBR layers to achieve high reflectivity for specific wavelengths while allowing other wavelengths to pass through. During optimization, three distinct reward functions were defined to enhance the optical uniformity of the micro-LED backlight module. Reward function 1 calculated rewards based on the difference between new and previous uniformity values, whereas reward functions 2 and 3 used fixed reference uniformity values of 75 and 79, respectively, with differences amplified by a cubic factor. The micro-LED backlight module with a DBR structure demonstrated a 32% improvement in uniformity performance compared with modules without a DBR. In addition, the DRL-based design method reduced computational time to just 17.9% of that required by traditional optical simulations. This study highlighted the potential of deep reinforcement learning in addressing complex optical design challenges and provided new directions for the development of future micro-LED display technologies, showing significant promise for applications in vertical-cavity surface-emitting lasers (VCSELs) and resonant cavity LEDs (RCLEDs).

In the same year, Piprek et al.44 conducted an in-depth study on the application of machine learning in optimizing the design of gallium nitride (GaN)-based LEDs. They evaluated practical cases of machine learning applications in LED design and proposed strategies to improve prediction accuracy. The study noted that although both machine learning and numerical simulation were effective methods for analyzing real-world phenomena, they carried inherent uncertainties and did not always successfully translate theoretical findings into practical applications when used together. Using genetic algorithm optimization, the research team identified the optimal doping profile to enhance LED efficiency after 500 generations of iterations, significantly reducing electron leakage from the quantum wells. In addition, they refined the electron blocking layer (EBL) into 10 layers of 3-nm-thick components with varying compositions, further mitigating leakage. This work demonstrated the potential of machine learning in improving LED performance and reliability evaluation while highlighting the challenges of applying these technologies in real-world production. The results showed that the optimized LED design achieved an increase in internal quantum efficiency (IQE) from 0.43 to 0.59 at a current density of 200 A/cm², emphasizing precise simulations’ importance and the need to consider material parameter uncertainties.

Both studies exemplify how advanced algorithms can optimize LED designs, with Huang et al. focusing on micro-LED backlight modules and Piprek et al. on GaN-based LEDs, collectively demonstrating the significant potential of machine learning and deep reinforcement learning in advancing LED technology. As shown in Fig. 2, the schematic diagrams of the micro-LED backlight module and the workflow of the environmental control agent illustrate the DBR structure’s role in enhancing optical performance, whereas the flow chart of the entire prediction framework based on machine learning underscores the importance of feature importance analysis in connecting model predictions to underlying physical mechanisms.

In 2023, Jiang et al.45 proposed a machine learning approach for optimizing the epitaxial structure of gallium nitride (GaN)-based LEDs to achieve full-color displays. This method utilized a dataset of GaN-based LED structures collected over the past decade and trained four typical machine learning models. Among these, convolutional neural networks (CNNs) provided the most accurate predictions for IQE and light output power density, achieving a root mean square error (RMSE) of 1.03% for IQE. The research team conducted a feature importance analysis to identify the key factors influencing LED performance. Using model interpretability methods such as permutation importance (PI) and SHAP value analysis, they connected model predictions to underlying physical mechanisms. Furthermore, the study demonstrated high-throughput screening capabilities, predicting the properties of over 20,000 structures within seconds to identify efficient LED designs. This machine learning-based design approach not only guided the selection of critical parameters in LED structure optimization but also significantly accelerated the development cycle of GaN-based LEDs. Looking ahead, this method may extend beyond optimizing existing LED structures to discovering novel semiconductor materials, innovating nanoscale manufacturing processes, and advancing quantum computing and communication technologies, further driving the semiconductor industry toward greater efficiency and intelligence. As shown in Fig. 3, the flow chart of the entire prediction framework based on machine learning and the schematic diagram of the mini-LED backlight module illustrate the critical role of feature importance analysis in connecting model predictions to underlying physical mechanisms and the effectiveness of AI algorithms in optimizing display technologies.

In 2023, Lee et al.46 introduced an innovative approach to optimizing mini-LED backlight module design by integrating artificial intelligence algorithms with optical simulation software. This study focused on improving the lighting uniformity of mini-LED structures, particularly by optimizing the thickness and concentration of the ${\mathrm{TiO}}_2$ scattering layer. The research team utilized LightTools optical simulation software in conjunction with MATLAB and employed a double deep Q-network (DDQN) algorithm as a reinforcement learning agent. This method significantly reduced the number of simulations, decreasing computational workload by 76.7% compared with traditional approaches. The DDQN algorithm leveraged two independent neural networks: an online network and a target network. The online network determined the next action, whereas the target network evaluated the Q-value of that action, mitigating overestimation issues. Using this intelligent optimization method, the team achieved an impressive 83.8% lighting uniformity, showcasing the advantages of incorporating advanced algorithms into display technology optimization. This work not only advanced mini-LED display technology but also provided new methodologies and insights for future display innovations. One can envision mini-LED backlight modules optimized through machine learning being integrated into wearable devices, offering users unprecedented interactive experiences. Moreover, these modules could revolutionize virtual and augmented reality, creating immersive visual experiences. As technology continues to evolve, these possibilities are steadily transitioning from science fiction to reality. Figure 4 illustrates the mini-LED backlight module and single packaged LED structure, the DDQN algorithm workflow, the optical simulation and algorithm design, the LightTools modeling process, and the simulation results comparison before and after DDQN algorithm optimization.

Deep ultraviolet light-emitting diodes (DUV LEDs), combined with quantum dots, hold significant potential for achieving full-color displays, representing a crucial advancement in the display field.47 In 2023, Feng et al.48 demonstrated an optimization of AlGaN-based DUV LEDs by employing an EBL structure based on InAlGaN/AlGaN superlattices and refining its design using the JAYA intelligent algorithm. Compared with conventional single-layer or double-layer EBL structures, the superlattice EBL effectively suppressed electron leakage and enhanced hole injection. This optimization resulted in a 41.2% increase in IQE at a current injection of 200 mA compared with single-layer EBL structures. Furthermore, the optimized EBL structure significantly mitigated efficiency droop at high currents, achieving a peak IQE of 0.48. This provides a novel approach to improving the efficiency of AlGaN-based DUV LEDs. Looking ahead, the integration of DUV LEDs with quantum dot technology could enable more precise color control and broader color gamut coverage, offering users an unparalleled visual experience. In addition, the potential of DUV LEDs in miniaturization and high-bandwidth communication could drive the development of innovative display technologies such as 3D and holographic displays. These technologies promise more immersive and interactive viewing experiences, as shown in Table 2, which pushes the boundaries of display innovation.

The application of AI technology in LED epitaxial design demonstrates remarkable potential for innovation. Algorithm-driven, high-precision simulations enable precise optimization of LED structures, enhancing performance and efficiency to facilitate the production of high-efficiency, high-stability full-color display technologies. This approach expands the application scope of LED devices in full-color displays and advanced display systems, addressing the market demand for high-performance display solutions. These cutting-edge studies not only accelerate the LED design process but also provide robust technological support for achieving richer color expression and higher-resolution display effects, signaling the future direction of innovation in full-color display technologies.

In the synthesis and coating processes of perovskite quantum dots, traditional experimental methods rely heavily on manual operations and empirical judgment.49 These approaches typically involve cumbersome chemical synthesis steps, multiple parameter adjustments, and repeated experimental validations. Such methods require substantial quantities of chemical reagents and solvents, making the process time-consuming, costly, and potentially harmful to the environment.50 Moreover, traditional techniques suffer from poor reproducibility and controllability, resulting in batch-to-batch variations in quantum dot quality, which hampers their scalability and limits their application potential in optoelectronic devices.51 These limitations underscore the urgent need for a more efficient, environmentally friendly, and cost-effective synthesis strategy. However, with the continuous advancement of artificial intelligence, a new optimization strategy has emerged.35 By employing machine learning algorithms to intelligently optimize the synthesis process of perovskite quantum dots, this approach significantly enhances the efficiency and quality of material synthesis, paving the way for broader and more reliable applications.52

Two-dimensional silver/bismuth organic-inorganic hybrid perovskites are materials with unique electronic and optical properties, combining the advantages of organic molecules and inorganic perovskite structures.53 These materials typically consist of organic amine cations, inorganic silver or bismuth anions, and halide anions, forming layered structures with a perovskite configuration.54^,55 Due to their exceptional optoelectronic properties, they exhibit high color purity, high luminescence efficiency, and wide spectral tunability, making them promising candidates for achieving high-resolution and high-contrast display technologies.56 In 2023, Wu et al.36 proposed a general framework integrating AI technologies to guide the experimental synthesis of materials, specifically addressing the synthesis challenges of 2D silver/bismuth organic–inorganic hybrid perovskites. This framework combines high-throughput experiments, chemical prior knowledge, and AI techniques, such as subgroup discovery and support vector machines, to efficiently screen a vast chemical space for perovskite materials with high synthetic feasibility. Figure 5 presents the framework for screening two-dimensional silver/bismuth iodide perovskites, detailing the problem-specific descriptors used in predicting synthesis feasibility. It shows the ROC curve and confusion matrix for the SVC model and uses SHAP values to illustrate the contribution of different descriptors to the prediction results. The figure also displays the influence of chemical groups on synthesis feasibility and provides information on material commercial availability, along with images of several synthesized perovskites. The research team developed an information feature set to quantify the stereochemical and topological properties of organic precursors. Using subgroup discovery methods, they identified regions in the chemical space favorable for forming 2D silver/bismuth iodide perovskites. Leveraging machine learning, they predicted 344 potential materials capable of forming 2D silver/bismuth perovskites from a pool of 8406 organic spacers. SHAP analysis further highlighted the importance of molecular topology in the formation of perovskites. Experimental validation successfully synthesized eight out of 13 predicted materials, demonstrating high synthetic feasibility for 2D silver/bismuth iodide perovskites. This study not only underscores the powerful potential of machine learning in optimizing material synthesis but also provides an effective strategy for the rapid discovery and understanding of novel functional materials.

Perovskite light-emitting diodes (PeLEDs) have garnered significant attention due to their high color purity, efficiency, and brightness, offering immense potential in display and lighting applications.57 In 2023, Lampe et al.58 introduced an innovative machine learning algorithm fusion method designed to rapidly optimize the synthesis process of perovskite nanocrystals. By integrating Gaussian processes, neural networks, and random forest classifiers, this approach significantly improved data efficiency, requiring only limited experimental data to optimize the synthesis of perovskite nanoplatelets (NPLs). Using this algorithm, the research team predicted the photoluminescence emission wavelength and spectral quality of ${\mathrm{CsPbBr}}_3$-based NPLs, using only the ratios of three precursors as input parameters. This strategy enabled the successful synthesis of previously challenging seven-layer and eight-layer thick NPLs and significantly enhanced the uniformity of two- to six-layer NPLs, as evidenced by narrower and more symmetrical photoluminescence spectra. The algorithm demonstrated high versatility, allowing the inclusion of additional synthesis parameters, making it applicable to less-explored nanocrystal syntheses. This not only improved synthesis quality but also reduced the time and resources needed to identify optimal conditions. As this research progresses, it promises revolutionary breakthroughs in nanocrystal synthesis. Machine learning algorithms will not only optimize the synthesis of perovskite nanocrystals but also facilitate the discovery of novel optoelectronic materials with extraordinary stability and efficiency, ushering in a new era of sustainable energy driven by intelligent materials. Imagine transparent solar windows, luminescent coatings, and self-healing smart devices becoming part of everyday life—these possibilities are rapidly transitioning from concept to reality.

In 2024, Zhao et al.59 developed an advanced machine learning model specifically designed to optimize the synthesis of perovskite quantum dots. By applying feature mask (FM) technology and a squeeze-and-excitation residual network (SEResNet) model, the team accurately identified critical parameters influencing quantum dot performance using limited sample data. Leveraging these models, they achieved precise control over the size, shape, and composition of perovskite quantum dots, optimizing their optoelectronic properties. In addition, the researchers implemented a parameter-tuning strategy based on a PI algorithm, which dynamically adjusted synthesis conditions according to the key features identified by the model. The FM-SEResNet model demonstrated high accuracy on the test set, achieving an RMSE of 0.833% and a correlation coefficient ($r$) of 0.980. The PI algorithm highlighted Pb, I, and MA as the key elements affecting power conversion efficiency (PCE). Experimental validation confirmed the model’s predictions, revealing that reducing the MA ratio improved PCE. Specifically, when the perovskite composition was ${\mathrm{MA}}_{0.05}{\mathrm{FA}}_{0.95}{\mathrm{PbI}}_{2.85}{\mathrm{Br}}_{0.15}$, the PCE reached a maximum of 21.78%, consistent with the model’s predictions (RMSE of 0.762%, $r=0.973$). This innovative approach not only enhanced the efficiency of quantum dot synthesis but also ensured their high performance in optoelectronic applications. Figure 6 presents the SEResNet model enhanced by the feature mask (FM) method for optimizing PSC performance prediction, covering PCE, VOC, JSC, and FF. It outlines the PI algorithm schema, from material properties to experimental validation. The figure also highlights key factors affecting PCE and compares experimental and predicted data trends for six PSC samples. As machine learning models continue to evolve, fully automated quantum dot manufacturing processes are anticipated, allowing real-time adjustments to size, shape, and composition for specific applications. Such intelligent synthesis platforms will significantly reduce R&D costs, shorten the time-to-market for novel optoelectronic devices, and enable personalized solar energy systems for households, heralding a new era of green energy solutions.

In 2023, Lukas et al.60 proposed a method combining deep learning and explainable artificial intelligence (XAI) to optimize the synthesis of perovskite nanocrystals. A major challenge addressed was the uncontrollable variability in quality during the large-scale processing of perovskite semiconductor thin films. By analyzing photoluminescence (PL) video data, the research team identified key features in the thin-film formation process, including drying, nucleation, crystal growth, and surface morphology development. The study revealed that higher PL peaks during the nucleation stage were associated with higher-quality perovskite thin films, whereas rapid PL signal decay during the crystal growth stage correlated with better performance. By optimizing solvent extraction rates and evaporation times during film formation, the uniformity and optoelectronic properties of the films were significantly improved. As XAI technology continues to advance, a comprehensive understanding of the perovskite solar cell fabrication process is becoming increasingly achievable. With precise predictions and optimizations enabled by deep learning models, the automated and intelligent production of solar cells will soon be realized. This will pave the way for more efficient, cost-effective, and environmentally friendly solar technologies, contributing significantly to the global transition towards sustainable energy solutions.

Organic molecular additives are critical for enhancing the performance of perovskite light-emitting diodes (PeLEDs) by improving crystal quality, reducing defects, and increasing luminous efficiency. However, due to limited data availability, the experimental screening of additives is costly and time-consuming, and traditional ML struggles to accurately predict their effectiveness.52 In 2022, Zhang et al.52 introduced a deep learning approach that successfully predicted the effectiveness of additives in PeLEDs with an accuracy of 96%. Using a small dataset of 132 molecules, they developed an efficient model named the enhanced molecular information model (EMIM). The EMIM workflow involves constructing a molecular dataset, converting molecular structures into machine-readable fingerprints and descriptors, training and integrating features using deep neural networks, and experimentally validating the model’s predictions. This approach maximized the utilization of molecular information while significantly reducing the redundancy commonly seen in traditional ML models for molecular screening. Through this innovative method, the research team achieved PeLEDs with an external quantum efficiency (EQE) of 22.7%, marking a new milestone in enhancing the performance of perovskite optoelectronic devices. This study opens a new pathway for leveraging deep learning to further improve PeLED performance and provides a powerful tool for advancing experimental research and applications in perovskite materials, as shown in Table 3. Table 3 details the AI-driven innovations in perovskite material synthesis, which underscores the transformative potential of our approach.

In summary, the application of AI in perovskite quantum dot synthesis significantly improves the efficiency and quality of optoelectronic device fabrication. This provides a robust foundation for widespread applications in energy, optoelectronic communication, and sensing.

In the process of preparing the color conversion layer, artificial intelligence technologies such as machine learning and deep learning can be used to intelligently and accurately regulate the parameters in the synthesis process, to improve the luminous efficiency and color purity of quantum dots at low cost and high efficiency. In addition, based on the analysis of a large number of experimental data, AI technology can refine specific ideas and methods in the synthesis and optimization process. With the advancement of deep learning and reinforcement learning algorithms, AI may be poised to enable real-time optimization of synthesis processes, automatically adjusting reaction conditions to achieve optimal quantum dot performance. Intelligent monitoring systems would even leverage computer vision to analyze quantum dot growth, predict and prevent production defects, and achieve ultra-high yield rates. Ultimately, AI technologies would drive the synthesis and application of perovskite quantum dots toward personalization and customization, meeting diverse demands ranging from flexible electronics to high-efficiency lighting.

In the early stages, defect detection in display panels primarily relied on manual visual inspection.61 Experienced operators were typically responsible for assessing nonuniform brightness and classifying products based on their observations.62 Although this method could identify obvious brightness irregularities to some extent, individual differences in visual acuity and sensitivity to brightness variations introduced subjectivity and inconsistencies in the evaluation process.63 For defect rectification, manufacturers generally addressed nonuniform brightness and related issues through manual adjustments, recalibration, or replacement of faulty components.64^–66 However, these methods were often inefficient and unable to guarantee a complete resolution of the problem in every instance.

In recent years, with the rapid advancement of AI technology, defect detection and automated repair have emerged as powerful tools for optimizing material and device fabrication processes.67^–69 By integrating high-precision image recognition and analysis technologies, these systems can accurately detect minute defects on or within materials in a short time and quickly devise repair strategies using machine learning algorithms. This intelligent detection and repair workflow not only enables efficient defect management but also provides customized solutions based on the specific properties of the materials, ensuring optimal performance. AI-integrated defect detection systems enhance identification accuracy and offer precise guidance for automated repair pathways, ensuring both efficiency and accuracy throughout the repair process.

Machine vision technology, through the application of computer vision and image processing algorithms, has enabled automated defect detection in industrial manufacturing, improving production efficiency and product quality.70 By minimizing human interference, it ensures consistency and real-time detection, reducing defect rates and production costs.67 In addition, the flexibility and data recording capabilities of machine vision systems provide robust support for product quality management and process optimization.68 In 2020, Zhang et al.71 proposed an innovative image processing method aimed at improving the accuracy and efficiency of nonuniform brightness defect detection. This approach cleverly integrated Gabor filters with Retinex theory to enhance the visualization of target defects in images. The Gabor filter effectively eliminated background noise and enhances local contrast, making subtle nonuniform brightness regions more prominent. Furthermore, Retinex-based global brightness equalization significantly increased the contrast between nonuniform brightness areas and surrounding normal regions by adjusting the overall brightness of the image. To meet the demands of industrial production for high efficiency, the researchers carefully selected optimal down-sampling parameters. This ensured the preservation of critical defect information while optimizing the algorithm’s computational speed. The method was tested on sample images collected from industrial sites, achieving a detection accuracy of 93.6% with an average detection time of only 0.632 s, meeting the requirements for industrial applications. This integrated approach, combining multiple image processing techniques, not only improves detection accuracy but also enhances the algorithm’s adaptability to real-world industrial environments. Looking ahead, the application prospects of this technology may extend to broader fields. For instance, incorporating multi-angle imaging and developing new algorithms could further improve detection versatility and precision.

Defect detection for displays using machine vision ensures objectivity and efficiency but lacks generality, requiring customized solutions for different screens and defect types, which increases costs. In addition, the demand for high-quality optical equipment limits its widespread adoption, and further improvements in detection accuracy are necessary.66 By contrast, deep learning-based defect detection plays a crucial role in industrial production, offering numerous advantages.72 By learning complex data features, it enhances accuracy, enables automated detection, reduces labor costs, and adapts to various defect types.34 It supports continuous operation, minimizes downtime, improves efficiency, and provides parallel processing capabilities suitable for large-scale production environments.25 In 2018, Yang et al.73 reported an innovative and efficient method called OSC-TL, which leveraged the representation capabilities of deep neural networks and the online sequential extreme learning machine (OS-ELM) strategy to achieve continuous online learning and classification of nonuniform brightness defects. This method demonstrated high-speed performance during both training and inference phases and accurately identified six distinct types of nonuniform brightness defects. The OSC-TL method exhibited remarkable efficiency, learning and recognizing Mura defect images within 1.5 ms with a detection accuracy of 93.6%, significantly outperforming other common classification algorithms. Moreover, it consumed far fewer computational resources and time than alternative approaches. Importantly, OSC-TL also enabled classifier training with a limited initial dataset, facilitating subsequent training on large-scale datasets and iterative updates for online networks. This makes it particularly well-suited for real-time processing applications.

Display nonuniformity (Mura) refers to visual irregularities on display panels, manifesting as inconsistent brightness that creates streaks or spots. This significantly degrades display quality and results in an unpleasant viewing experience.74 Mura appears in various forms, with banding Mura and linear Mura being the most common. Banding Mura presents as vertical or horizontal brightness differences, often caused by material inconsistencies or process defects. Linear Mura, on the other hand, appears as high-contrast lines, typically associated with physical damage or circuit issues.34 Due to the irregular shapes, varying sizes, and low contrast of Mura, traditional detection methods struggle to accurately identify them, particularly against nonuniform backgrounds, making Mura detection a challenging aspect of display quality control.75 In 2021, Xie et al.34 proposed a novel Mura detection method based on an improved generative adversarial network (GAN), named UADD-GAN. This method specifically addressed the difficulty of identifying Mura in quality inspection processes for display panels. The researchers trained the generator using only normal samples, enabling it to model the distribution of normal data. During detection, the generator poorly reconstructed samples containing Mura, allowing for effective differentiation from normal samples. To fully leverage the discriminator, the researchers incorporated both classification layers and multiple feature layers from the discriminator to enhance the generator’s ability to reconstruct normal samples accurately. In addition, to handle nonsquare samples, a bilateral detection approach was implemented, significantly improving detection accuracy. Experimental results demonstrated that this method outperformed state-of-the-art techniques across various Mura datasets, achieving an average detection accuracy of 0.923 on banding Mura datasets and $\sim 0.933$ on linear Mura datasets. This provides an efficient and accurate solution for quality inspection in display panels.

In 2021, Kidoguchi et al.76 proposed an innovative quantitative solution to address the long-standing challenge of detecting and evaluating display non-uniformity (Mura) in display quality control. This method enabled precise quantitative analysis of brightness and color irregularities by leveraging a deep convolutional neural network (DCNN), specifically a convolutional autoencoder (CAE) featuring four convolutional layers. Although previous studies have proposed various DCNN models for anomaly detection, these methods often struggled to handle multiple Mura defects on a single screen, resulting in limited correlation with human visual assessments. To overcome these limitations, the research team adopted the CAE, an unsupervised machine learning technique trained exclusively on defect-free data. This approach optimized network parameters to accurately reconstruct normal data. During the reconstruction of anomalous data, the CAE exhibited inaccuracies due to its lack of exposure to such data, a characteristic effectively utilized for anomaly detection. The researchers further integrated the contrast sensitivity function (CSF), a well-known model of human visual spatial frequency characteristics, to emphasize various Mura defects before inputting the data into the CAE. This enhancement significantly improved the accuracy of automated detection systems. The study not only advances the precision of Mura detection but also opens avenues for predicting and automatically correcting potential display anomalies, pushing display quality to unprecedented levels of perfection and delivering flawless, immersive visual experiences to users.

In 2023, Lee et al.33 developed an advanced artificial intelligence–assisted automatic optical inspection (AOI) system designed to detect stains during the backplane manufacturing stage of OLED panels. By combining explainable artificial intelligence (XAI) predictive models with test element group (TEG) engineering methods, the system significantly improved Mura detection accuracy. The research team conducted an in-depth analysis of the dehydrogenation mechanism during the backplane manufacturing process and discovered that hydrogen doped into the driving transistor (TR) near contact holes undergoes dehydrogenation due to high energy during the source-drain film formation process. Notably, the distance between the TR and the contact holes was found to affect TR characteristics, thereby influencing panel brightness uniformity. As this distance increases, the ion of the TR decreases likely due to the diminished impact of dehydrogenation on TR performance, which reduces Ion levels. By analyzing these data, the AOI system could predict potential TR performance variations under different manufacturing conditions, allowing early identification of areas likely to cause brightness nonuniformity. This capability facilitates real-time adjustments to process parameters during production, optimizing TR performance and ensuring brightness uniformity in OLED panels. In addition, the study introduced a proportional–integral–derivative (PID) control strategy based on logistic regression. By leveraging the proportional (P), integral (I), and derivative (D) components, the system dynamically adjusted the built-in parameters of display panels according to the identified defect types. This innovative approach not only enhanced defect correction efficiency but also achieved the goal of eliminating brightness nonuniformity in micro-LED displays during production, ensuring 100% uniformity. Figure 7 illustrates the dehydrogenation mechanism, the relationship between ion experimental values and the distance between contact holes and driving TR, and the process of parameter extraction from transistor TEG images for AI model training. It also displays Mura simulation results using the image-quality prediction system.

In 2023, Chen et al.70 proposed an innovative lightweight YOLO-ADPAM detection method that significantly improved object detection performance by incorporating attention mechanisms. The researchers cleverly utilized the complete-intersection-over-union (CIoU) loss function to design a K-means-CIoU++ clustering algorithm, optimizing anchor box dimensions within the dataset. This approach not only enhanced the precision and stability of bounding box regression but also improved the overall recognition and localization accuracy of the algorithm. In addition, the research team introduced a parallel attention module (PAM) that integrates channel and spatial attention mechanisms. This module efficiently extracts critical feature information from feature maps while preserving essential spatial details. For the neck network design, they incorporated atrous spatial pyramid pooling (ASPP) and depthwise separable convolutions (DSCM), expanding the receptive field of the feature maps and significantly improving detection accuracy, especially for small objects. Experimental results in TFT-LCD defect detection demonstrated that the YOLO-ADPAM model achieved an impressive mean average precision of 98.20% and a detection speed of 83.23 frame/s, meeting the dual requirements of accuracy and speed for real-time detection tasks. Looking ahead, the YOLO-ADPAM model is expected to evolve into a holographic intelligent detection system integrated into nanodrones. This system could enable instant, contactless defect detection for TFT-LCD displays and even predict potential manufacturing defects. Such advancements could push detection precision to the atomic scale, revolutionizing the industrial visual inspection field. Figure 8 presents 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.

Looking ahead, the application of AI technology in defect detection and automated repair processes for display panels is poised to achieve significant breakthroughs. As AI technology continues to advance, algorithms are expected to gain a deeper understanding of complex image data, enabling more accurate identification and classification of various defects. Current AI-based detection systems already offer rapid, automated defect recognition, capable of categorizing detected issues and autonomously selecting appropriate repair solutions. This automation not only reduces labor demands but also enhances production efficiency and product quality. In the future, such systems are anticipated to further improve detection speeds, enabling near-real-time feedback, which will play a critical role in boosting production efficiency and reducing costs. As the technology matures and evolves, the display manufacturing industry is expected to become increasingly intelligent, better equipped to meet the market’s growing demand for high-quality display products.

Using the AI technologies such as machine learning and deep learning, the image data of the display panel can be analyzed to automatically identify and classify various defects, for example, using CNN to process the acquired panel images could quickly and accurately detect problems such as uneven brightness and pixel defects. In addition, compared with traditional manual detection methods, AI could achieve more efficient and objective defect detection, thus greatly improving the detection speed and accuracy of defect detection. Moreover, AI could adaptively learn to detect new defect types with excellent versatility and scalability. In the future, AI systems will leverage advanced self-evolving algorithms combined with quantum computing capabilities to perform real-time, holographic 3D scanning of displays. These systems will precisely identify and classify even the smallest defects. Miniature robots equipped with integrated AI chips, deployed via drones, will autonomously navigate to defect locations and employ nanoscale repair techniques, such as laser welding and molecular reconstruction, to perform on-site corrections, ensuring the display quality reaches unparalleled perfection.53 In addition, AI systems will seamlessly integrate with manufacturing processes to predict potential defects and enable preventive maintenance, significantly enhancing production efficiency and product quality.77 This transformative innovation will usher in a new era of intelligent manufacturing, setting new benchmarks for technological excellence.

Local dimming is an advanced display technology that dynamically adjusts the brightness of specific backlight zones in LCDs to optimize image contrast and color performance. By precisely controlling backlight intensity based on the brightness variations in the displayed content, this technique significantly improves visual quality, particularly when showcasing high dynamic range (HDR) content. Compared with traditional global dimming, local dimming excels in preserving image details, preventing brightness loss in dark areas, and overexposure in bright areas. In addition, it reduces energy consumption, contributing to the enhanced environmental sustainability of displays. A specific implementation of local dimming is zone dimming, where the backlight is divided into multiple smaller zones, each capable of independent brightness control. This allows for targeted optimization of different parts of the image. For instance, if a particular area of the screen is dark, only the corresponding backlight zone dims, whereas the brightness in other zones remains unaffected. This approach improves image contrast and detail representation, delivering a more immersive and visually striking experience. Figure 9 illustrates the local dimming process, showing how it enhances image quality through targeted backlight adjustment.

In recent years, the integration of AI has revolutionized dimming technology. By leveraging deep learning algorithms, AI can analyze image content, intelligently identifying brightness distribution and detail features to enable more precise and adaptive dimming strategies. These algorithms dynamically adjust backlight brightness based on image content, optimizing display performance, enhancing contrast and color accuracy, and reducing power consumption. Moreover, AI effectively mitigates halo effects by recognizing and suppressing them through intelligent algorithms, improving visual comfort.78 In practical applications, AI dimming systems can automatically adjust display parameters based on user viewing habits and changes in ambient lighting, delivering a personalized and seamless viewing experience.

In the field of local dimming for LCDs, traditional algorithms often rely on manual feature extraction and real-time optimization, which may struggle to adapt to diverse image content and exhibit limitations when addressing varying display requirements. To overcome these challenges, researchers have turned to deep learning techniques to enhance the performance of local dimming in LCDs. In 2019, Song et al.37 proposed a deep learning-based pixel compensation algorithm for quantum dot backlit LCDs with local dimming. The algorithm employed a convolutional neural network called the local dimming neural network (LDNN), which directly generated compensated images from input images without requiring any information about backlight dimming levels. Built on the U-net architecture, LDNN used skip connections to retain high-resolution features in the upsampling path while leveraging bilinear interpolation to reduce parameter complexity without compromising image quality. The network was trained and validated on the DIV2K 2K image dataset. Experimental results demonstrated that the proposed deep learning algorithm significantly outperformed traditional weighted pixel compensation methods in terms of peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and color difference (CD). Figure 10 illustrates the architecture and effectiveness of the LDNN for local dimming in quantum dot backlit LCDs. This work represents the first application of deep learning to local dimming algorithms, offering new possibilities for optimizing LCD energy efficiency and enhancing image quality.

Light crosstalk, also known as light leakage or light diffusion, refers to the phenomenon in LCD backlight systems where light from one area affects adjacent areas. This is particularly noticeable in black regions, resulting in unwanted brightness that reduces image contrast and leads to loss of detail. This issue is especially pronounced in local dimming technology, which lowers backlight brightness in specific regions to enhance image contrast but can degrade the quality of neighboring regions if not properly managed. In 2020, Zhang et al.79 proposed a novel local dimming technology based on deep convolutional neural networks (Deep CNNs) to improve the backlight systems of LCDs. This approach incorporated two key components: the backlight luminance extraction network (BLEN) and the pixel compensation network (PCN). The aim was to enhance LCD contrast while reducing energy consumption. The research team first utilized seven traditional backlight luminance extraction algorithms and five conventional pixel compensation methods to generate high-quality backlight luminance and pixel-compensated images as benchmarks. They then trained and validated their proposed deep learning networks on a custom-built image dataset. Experimental results showed that the deep CNN-based local dimming technology outperformed traditional methods in contrast enhancement and detail preservation. When applied to LCD-LED dual-modulation displays, the proposed method achieved a contrast ratio (CR) of 9.26, significantly higher than the highest value of 8.89 obtained using the TSPCM method. In addition, the PSNR and SSIM values of 25.31 and 0.90, respectively, were comparable to or slightly better than those of the TSPCM method, demonstrating its superior image quality. This study not only successfully established two datasets as a foundation for developing local dimming networks but also highlighted the remarkable ability of CNNs to extract complex image features and reduce light crosstalk. These advancements mark a significant step forward in improving the performance of LCD backlight systems.

High dynamic range (HDR) display technology, compared with traditional standard dynamic range (SDR) display technology, offers a significantly broader range of brightness and color. By increasing the ratio between maximum and minimum brightness levels—known as dynamic range—HDR displays reveal greater detail in both highlights and shadows, resulting in enhanced image quality that appears more realistic and natural.12 In 2020, Duan et al.80 proposed a novel deep learning-based local dimming method for rendering HDR images on dual-panel HDR displays. This approach utilized a CNN to directly predict the backlight values for each dimming zone based on the input HDR image. The model was designed and trained with adjustable power parameters, allowing users to balance power consumption and display quality according to their needs. The proposed method was evaluated against six other techniques using a test set of 105 HDR images and various quantitative quality metrics. The results showed that the method outperformed all alternatives in both display quality and power efficiency. Specifically, the HDR images rendered using this method achieved higher fidelity in terms of brightness and color reproduction compared to traditional SDR displays while maintaining a dynamic range of up to 50,000:1 and peak brightness levels of $\sim 8500\mathrm{cd}/{\mathrm{m}}^2$. It paves the way for further advancements and applications in HDR display technology.

Building on the progress in local dimming technology for LCDs in 2020, Duan et al.81 introduced an innovative deep learning-based local dimming method, deep controllable backlight dimming (DBLD), in 2022. Designed specifically for rendering HDR images on dual-panel HDR displays, DBLD employed a CNN to predict backlight values directly from HDR images. By incorporating a controllable power parameter, the method enabled an optimal balance between power consumption and display quality. Evaluations conducted on a test set of 105 HDR images demonstrated that DBLD outperformed existing technologies in both display quality and power efficiency. Within the power parameter range of 0.6 to 0.9, the power-saving ratio (PSR) increased without significant changes to PU-PSNR (53.75) and PU-MS-SSIM (0.99), whereas HDR-VDP-2.2 values exhibited a gradual decline. For power parameters between 0.9 and 1.1, PU-MS-SSIM remained at 0.99, but PU-PSNR and HDR-VDP-2.2 saw noticeable decreases. Beyond 1.1, particularly in the 1.1 to 1.5 range, PU-PSNR and HDR-VDP-2.2 declined significantly, and PU-MS-SSIM dropped sharply after a slight decline from 1.1 to 1.4. DBLD’s controllable power parameter design allows users to trade off power consumption against display quality, offering a tailored balance for various use cases. This deep learning-based local dimming technology is anticipated to achieve commercial adoption across a range of consumer electronics, including smartphones, tablets, and televisions. In the future, HDR displays powered by DBLD could intelligently adjust based on user preferences and ambient lighting conditions, delivering customized viewing experiences while maximizing energy efficiency.

In 2024, Liu et al.82 proposed a deep learning-driven pixel compensation method to optimize the display quality and reduce energy consumption in liquid crystal display (LCD-LED) systems. The algorithm was based on the U-Net architecture, incorporating both downsampling and upsampling processes. Skip connections were used to combine deep features extracted during downsampling with the feature maps from upsampling, accelerating network convergence, and mitigating the degradation issues associated with increased network depth. By adjusting the brightness of the backlight matrix, the algorithm compensated for the loss of pixel details due to dimming, enhancing contrast and reducing power consumption. To address optical crosstalk, the study simulated the light diffusion process from the backlight panel to the display panel, ensuring that the backlight matrix resolution matches the display image resolution, thereby minimizing image distortion caused by crosstalk. Experimental results demonstrated that the proposed algorithm outperformed traditional methods in terms PSNR and SSIM. Specifically, compared with traditional nonlinear, log function compensation (LFC), and linear compensation methods, the proposed algorithm achieved scores of 266.65, 23.13, 0.998, and 0.23 for color ratio, PSNR, SSIM, and color deviation, respectively, outperforming nonlinear methods (188.96, 21.24, 0.996, and 0.28), LFC methods (261.74, 9.61, 0.972, and 0.32), and linear methods (225.20, 14.78, 0.988, and 0.75). This research combined the advantages of different pixel compensation algorithms, demonstrating excellent performance in reducing backlight power consumption and enhancing static contrast while preserving more original image details. It holds promise for applications in VR and AR, offering users a more immersive visual experience.

In conclusion, the application of AI technology in full-color display dimming not only enhances the user’s viewing experience but also helps reduce energy consumption, promoting energy efficiency and environmental sustainability. With the continuous development of technology, it is expected that future display devices will achieve more comprehensive breakthroughs in flexibility, contrast, color performance, and energy efficiency, further improving the user experience. For instance, through real-time analysis of image content to optimize the brightness and color of each pixel, higher energy efficiency can be achieved without sacrificing image details. In the fields of VR and AR, AI-driven dimming technology will provide a more realistic and immersive experience.83 As wearable devices become lighter and thinner, AI technology will help these devices deliver visual effects indistinguishable from the real world while also reducing eye fatigue. Ultimately, AI technology will enable intelligent, personalized, and eco-friendly display solutions in the full-color display dimming field, driving display technology toward greater efficiency, superior experiences, and improved environmental sustainability.

This study reviews the applications of AI technology in full-color display systems, focusing on epitaxial structure design, defect detection and repair, perovskite synthesis and coating, and dynamic dimming technologies. With continuous technological advancements and increasing industrial demands, AI’s powerful data processing and pattern recognition capabilities have brought new vitality to the development of full-color displays. In epitaxial structure design, AI-driven deep learning models enable precise prediction and optimization of LED structural parameters, significantly enhancing design efficiency and accuracy. For defect detection and repair, AI improves detection precision and repair efficiency, ensuring high-quality production of display panels. In perovskite synthesis, AI optimizes synthesis processes through machine learning algorithms, enhancing material quality and enabling precise fabrication of multidimensional perovskite patterns. Furthermore, in dynamic dimming, AI introduces adaptive strategies for precise adjustments, improving display performance and energy efficiency.

From an industrial perspective, the integration of AI into full-color display systems offers significant benefits. For example, AI-powered defect detection reduces human error and ensures high-quality production of display panels. AI-driven predictive maintenance models use real-time data from sensors to detect potential failure points in production equipment, reducing downtime by 10% to 40% and increasing machine life by 5% to 10%. This capability is crucial for maintaining high production standards and minimizing disruptions. AI-driven optimization in perovskite synthesis can streamline manufacturing processes and enhance material efficiency. These advancements not only improve product quality but also reduce production costs and accelerate time-to-market. Moreover, AI-powered quality control systems, equipped with advanced vision systems, can detect defects in real time on production lines, ensuring that only high-quality products reach the market. This not only enhances product consistency but also reduces waste and rework costs. In addition, AI in energy management optimizes energy use by predicting demand, improving both sustainability and cost efficiency.

However, the integration of AI in full-color display systems also faces several challenges. For example, the computational power required to run complex AI models can be substantial, potentially leading to increased energy consumption and system costs. In addition, ensuring the efficiency and scalability of AI systems in industrial settings remains a critical issue as real-time processing and high throughput are essential for large-scale production. Furthermore, the cost of implementing AI technologies, including hardware upgrades and software development, may pose a barrier to widespread adoption.

Beyond these technical challenges, the integration of AI in emerging areas such as human–computer interaction (HCI) and brain–computer interfaces (BCI) presents additional complexities, thus related academic investigation is still necessary. For instance, in HCI, ensuring user-friendly and intuitive interfaces is crucial for the success of smart displays and VR/AR devices. Meanwhile, BCI technologies face challenges related to data privacy, user consent, and the potential loss of support for implanted devices. In the context of the metaverse, the seamless integration of AI with virtual and augmented reality platforms requires addressing issues related to user experience, data security, and ethical considerations.

Looking ahead, AI’s role in full-color display technologies is expected to deepen and expand, driving advancements toward higher efficiency, superior user experience, and smarter solutions. As AI matures, it is poised to revolutionize the display industry, reshaping our digital world and ushering in a new era of vibrant and intelligent visual experiences.