image: This schematic illustrates the integrated AI for Materials (AI4M) framework proposed for OLED material design. It starts from (I) Basic Luminescent Materials, focusing on structural and photophysical properties such as IQE, PLQY, FWHM, and stability that constructing optimizing goals for material design; it then highlights three major components involved in AI4M framework: (II) Quantum Chemistry Calculations, employing DFT and advanced post-HF methods to provide reliable molecular descriptors and datasets; (III) Prediction Models and Model Interpretation, using machine learning to establish property prediction models and identify key molecular features; and (IV) Molecular Generation and Screening, applying high-throughput screening and inverse design to efficiently explore and optimize chemical space.
Credit: ©Science China Press
Organic light-emitting diodes (OLEDs) have become a core technology for next-generation display and lighting applications, placing extremely high demands on emissive materials to achieve a delicate balance of high efficiency, long operational lifetime, and superior color purity. Such goals require breakthrough innovation at the molecular level. However, the conventional approach—largely dependent on empirical knowledge and limited molecular modifications—has shown clear limitations when faced with the exponentially growing chemical space and the complex, multidimensional optimization required for performance trade-offs. Against this backdrop, artificial intelligence (AI), especially when combined with quantum chemistry, machine learning, and generative models in the so-called AI4M (AI for Materials) framework, is widely seen as a promising way to overcome these bottlenecks. Yet, the unique challenges of OLED materials—such as multi-scale electronic coupling, intricate vibrational effects, and strong nonlinear structure–property relationships—mean that applying AI effectively in this field is rather challenging. To bridge this gap between AI applications and OLED material theory, the authors of this review propose a systematic AI-driven framework for OLED material design, offering both theoretical guidance and practical pathways for intelligent materials development.
The review first provides an overview of key metrics that define OLED device performance and their current development status, including internal quantum efficiency (IQE), photoluminescence quantum yield (PLQY), emission wavelength, full width at half maximum (FWHM), color purity, and device lifetime. It highlights that these performance indicators are fundamentally governed by factors such as molecular orbital distribution, excited-state behavior, and vibrational coupling at the molecular scale. These parameters often impose competing requirements—for example, in blue narrowband emitters, achieving the right balance among emission wavelength tuning, molecular backbone rigidity, and synthetic feasibility remains a core challenge. The paper further analyzes the structural features and emission mechanisms of fluorescence, phosphorescence, TADF, and MR-TADF materials, comparing how these different systems perform in terms of efficiency, stability, and color purity.
The review underscores the fundamental role of quantum chemistry calculations within the AI4M framework, serving as a crucial bridge between microscopic molecular structures and macroscopic material properties. Quantum chemical data not only provide reliable labels for training machine learning models but also ensure their physical consistency and interpretability. The authors compare different functionals (LDA, GGA, hybrid, double-hybrid) for predicting parameters such as HOMO-LUMO levels, dipole moments, and electrostatic potentials, noting that method selection has a significant impact on consistency, especially for molecules involving long-range charge transfer (CT) states. Balancing accuracy with the computational efficiency needed for large-scale dataset generation remains a critical challenge. In excited-state calculations, the review analyzes the systematic bias of TD-DFT in predicting CT-state energies and discusses improvements such as TDA approximations, post-Hartree–Fock methods, and the importance of solvent models like LR-PCM and SS-PCM for adjusting excitation energies and transition dipole moments. The authors also emphasize that parameters such as reorganization energy, Franck–Condon factors, Huang–Rhys factors, and oscillator strengths—closely linked to emission efficiency and spectral features—are vital considerations when constructing usable datasets.
The section on machine learning models for property prediction clearly lays out the end-to-end workflow for applying ML to OLED materials — covering dataset preparation, feature engineering, model development, and interpretability analysis. The authors emphasize that high-quality, diverse, and well-curated training datasets form the foundation of any reliable AI prediction model. However, they point out that challenges such as uneven data distributions and the risk of overfitting with small datasets still need to be addressed. The authors emphasize that machine learning serves as a critical bridge linking molecular features to target properties, with different algorithms offering advantages depending on sample size and problem complexity. They also stress that interpretability is essential when applying models to OLED design. Tools like SHAP and LIME can help researchers quantify how each molecular feature affects prediction results, pinpoint key structural fragments, and verify structure–property relationships—helping prevent models from becoming black boxes and ensuring AI genuinely serves design and mechanistic discovery. For OLED applications, the review summarizes progress in predicting core optoelectronic properties, from orbital energy levels and emission wavelengths to PLQY and FWHM, covering TADF, MR-TADF, and Ir(III) complexes. By combining experimental data, quantum calculations, and deep learning, more models now enable rapid, accurate property prediction while providing structural insights that guide rational design.
Despite this promise, the authors point out limitations such as poor transferability, narrow sample spaces, single-property prediction, and neglect of intermolecular interactions. Factors like exciton diffusion, triplet quenching, and molecular orientation require multi-scale simulation and multi-source feature inputs for accurate description. In the future, integrating multi-objective optimization, physical constraints, and experimental feedback will be key to building fully explainable, generalizable AI models that can span the full chain from molecules to devices.
The section on AI-guided material screening and design introduces different AI-driven molecular design strategies, including high-throughput screening, molecular generation, and inverse design. High-throughput screening uses a funnel-type hierarchical workflow to progressively filter and optimize large chemical spaces. The authors note that constructing a diverse chemical space is the first step, followed by cheminformatics-based rule filtering, rapid ML predictions, quantum chemical validation, and synthetic feasibility assessment—ultimately forming a closed-loop screening system with experimental feedback. Generative models play a crucial role in building diverse chemical spaces. The paper reviews recent breakthroughs in deep learning–driven molecular generation for small molecules, highlighting its inspiration for OLED design. It systematically explains key generative algorithms like RNNs, GANs, and VAEs, and compares how LSTM, GRU, and new Transformer architectures improve molecule validity, diversity, and distribution matching. Unlike forward design and high-throughput screening, inverse design works backward from target properties to molecular structures, offering a new paradigm for developing organic emissive materials. Techniques such as Bayesian optimization, reinforcement learning, and genetic algorithms can be combined with generative models and multi-objective optimization to accurately explore high-potential regions.
Building on this, the authors provide real-world examples of AI-driven OLED material design, highlighting TADF, phosphorescent, and MR-TADF cases that demonstrate closed-loop pathways combining fragment assembly with multiple rounds of HTVS and experimental validation. For instance, Gómez-Bombarelli et al. constructed a virtual library of 1.6 million TADF candidates and used neural networks plus quantum calculations to identify high-performance molecules with EQE up to 22%. Similar strategies have since been successfully applied to Ir(III) complexes and HTL/ETL materials, proving the paradigm's feasibility and scalability. While AI-driven OLED molecule generation and screening have made significant strides, the paper notes persistent challenges, including data scarcity and generative model limitations for handling complex polycyclic structures, heteroatoms, and coordination bonds. The authors also point out that synthetic feasibility assessments are still too simplistic and that deeper integration with retrosynthesis tools and standard benchmarking systems will be crucial for future development.
While AI-driven OLED molecule generation and screening have made significant strides, the paper notes persistent challenges, including data scarcity and generative model limitations for handling complex polycyclic structures, heteroatoms, and coordination bonds. The authors also point out that synthetic feasibility assessments are still too simplistic and that deeper integration with retrosynthesis tools and standard benchmarking systems will be crucial for future development.
The review concludes that for AI to deliver real impact in OLED material research, several key advances are needed. First, expanding high-quality, multidimensional experimental datasets covering different electronic states and performance parameters will be essential. Second, integrating quantum chemistry calculations with machine learning across scales will enable accurate prediction and mapping from molecules to device performance. Third, specialized generative models tailored for the unique emission characteristics of OLEDs—especially for narrowband emission, molecular stability, and synthetic feasibility—must be developed to overcome current model limitations. Finally, the authors emphasize the importance of broader, deeper interdisciplinary collaboration to build efficient, reusable platforms for closed-loop materials design and validation.
In summary, this review systematically maps the current landscape of AI in OLED research, providing detailed analysis of quantum chemistry, machine learning algorithms, and generative models for OLED design. Given the relatively limited work at this intersection so far, the paper draws on successful AI applications in related materials systems, offering valuable cross-domain insights for advancing intelligent OLED development. The proposed framework serves as a comprehensive toolkit and practical roadmap for researchers while clarifying key technical bottlenecks and future directions for this rapidly evolving field.