Abstract: Photoactive iridium complexes are of broad interest due to their applications ranging from lighting to photocatalysis. However, the excited state property prediction of these complexes challenges ab initio methods such as time-dependent density functional theory (TDDFT) both from an accuracy and a computational cost perspective, complicating high throughput virtual screening (HTVS). We instead leverage low-cost machine learning (ML) models to predict the excited state properties of photoactive iridium complexes. We use experimental data of 1,380 iridium complexes to train and evaluate the ML models and identify the best-performing and most transferable models to be those trained on electronic structure features from low-cost density functional theory tight binding calculations. Using these models, we predict the three excited state properties considered, mean emission energy of phosphorescence, excited state lifetime, and emission spectral integral, with accuracy competitive with or superseding TDDFT. We conduct feature importance analysis to identify which iridium complex attributes govern excited state properties and we validate these trends with explicit examples. As a demonstration of how our ML models can be used for HTVS and the acceleration of chemical discovery, we curate a set of novel hypothetical iridium complexes and identify promising ligands for the design of new phosphors.
Introduction
Iridium(III) complexes serve as the "workhorses" of modern technology, particularly in Organic Light-Emitting Diodes (OLEDs) and green photocatalysis, due to their ability to efficiently convert electrical or chemical energy into light. Despite their importance, the design of new phosphors remains a slow and resource-intensive process. Traditionally, researchers have relied on Time-Dependent Density Functional Theory (TDDFT) to predict molecular properties. However, TDDFT faces significant hurdles: a single calculation can exceed 24 hours, and it often struggles with the complex metal-organic interactions inherent in these systems, frequently overestimating emission energies. These limitations effectively prohibit High-Throughput Virtual Screening (HTVS), where thousands of candidates must be evaluated rapidly.
Methodology: Bridging Physics and Machine Learning
To overcome these bottlenecks, we developed a machine learning framework capable of predicting excited-state properties with high speed and accuracy. A central challenge in applying ML to chemistry is determining how to represent a molecule to an Artificial Neural Network (ANN). We evaluated several descriptors, including structural fingerprints (Morgan and Dice) that describe atom connectivity, graph descriptors (RACs), and electronic structure features calculated via extended Tight Binding (xTB).
The complexes studied generally follow the $[Ir(CN)_2(NN)]^+$ architecture, where the properties are tuned by varying the cyclometalating (CN) and ancillary (NN) ligands. While simple connectivity-based fingerprints perform well on data similar to the training set, they often fail to generalize to novel ligand types. We found that xTB features—a semi-empirical middle ground between simple physics and full quantum mechanics—provide the necessary physical context for the model to make accurate predictions on unfamiliar chemical space.
Results and Model Performance
The models were trained on a high-quality experimental dataset of 1,380 iridium(III) complexes. Training on experimental data is crucial as it allows the model to capture real-world complexities, such as non-radiative decay and solvent effects, that pure simulations might miss.
As shown in the performance analysis, models utilizing electronic features (xTB, B3LYP, $\omega$PBEh) maintained significantly lower error rates when tested on "grouped splits"—scenarios where the model must predict properties for ligands it has never encountered during training. Specifically, the xTB-based ANN achieved a Mean Absolute Error (MAE) of just 0.021 eV for emission energy, dramatically outperforming TDDFT, which typically exhibits an average overestimate of 0.3 eV. Furthermore, the ML approach offers a massive speedup, enabling the evaluation of thousands of complexes in seconds rather than days.
Chemical Insights and Discovery
Feature importance analysis revealed that the Ionization Potential (IP) of the CN ligand is the primary driver for emission energy; higher IP values correlate with higher emission energies (shifting light toward the blue end of the spectrum). Interestingly, our results suggest "orthogonal tuning" of properties: the CN ligand primarily dictates color and lifetime, while both CN and NN ligands influence brightness (spectral integral). This insight allows chemists to modify one ligand to adjust color without inadvertently compromising the phosphor's brightness.
To demonstrate the utility of this framework for discovery, we screened over 60,000 hypothetical complexes from the Cambridge Structural Database. By applying Uncertainty Quantification (UQ), we identified a curated set of ~3,600 promising candidates, including specific ligands predicted to yield phosphors with exceptional lifetimes and high brightness.
Conclusion
This low-cost ML approach provides a robust alternative to traditional ab initio methods for excited-state property prediction. By leveraging semi-empirical electronic features, we achieve high transferability and accuracy suitable for guiding experimental efforts. Future work will focus on extending this framework to other transition metals and integrating it into active learning loops to accelerate the discovery of sustainable, high-efficiency light emitters.
Summary updated