Multi-level Δ-learning for predicting photophysical properties of phosphorescent platinum(II) complexes

Abstract

The utilization of phosphorescent metal complexes as emissive dopants for organic light-emitting diodes (OLEDs) has been the subject of intense research. Pt(II) complexes are particularly popular triplet emitters due to their color-tunable emissions and ease of preparation. To make them viable for practical applications as efficient OLED emitters, it is essential to develop Pt(II) complexes with superior photophysical properties. However, as studying metal complexes is not as easy and straightforward as pure organic molecules, which is probably due to the coordination bond, low sample data and representation issues, very few research employs machine/deep learning (ML/DL) method to explore Pt-complex and achieves high efficiency and accuracy. To this end, first-principles calculations and ML/DL algorithms combined with Δ-learning approach are employed to predict the photophysical properties of interest. This offers a powerful way to evaluate Pt-complexes from different accuracy levels which can achieve high throughput virtual screening (HTVS) and highly accurate predictions even in small sample scenarios. With respect to the dataset, ca. 200 experimental results are collected from literature. The experimental dataset is then augmented with two new datasets including 526,000 unlabeled Pt-complexes structures and ca. 500 first-principles optimized structures with simulated results. Besides, three ML protocols are proposed. They achieve highly accurate ensemble ML predictions of photophysical properties using first-principles calculated features, high throughput virtual screening of potential Pt-complexes based on rapid-computed features, and multi-level Δ-learning for experimental radiative decay rate constant by semi-supervised GNN models, respectively. Through the three protocols, issues such as representations of Pt-complex, generations, low experimental samples, HTVS and accurate predictions are addressed. To demonstrate the capacity of the proposed protocols, all the ML models have shown good performance on both the independent testing set and the external testing set. These protocols are expected to serve as valuable tools for predicting photophysical properties of metal complexes, accelerating the rapid development of novel OLED materials. More importantly, the Δ-learning approach offers guidance for the future development of ML models for small-sample chemical systems, such as metal-complexes.