Multi-level Δ-learning for predicting the radiative decay rate constant of phosphorescent platinum(ii) complexes

Abstract

Phosphorescent metal complexes, especially Pt(II) complexes, are widely used as emissive materials in organic light-emitting diodes (OLEDs) owing to their tunable emissions and ease of preparation. To enable practical OLED applications, it is essential to develop Pt(II) complexes with a high radiative decay rate constant (k_r). To this end, an efficient and accurate prediction tool for a small experimental k_r sample is highly desirable. In this work, we propose a multi-level Δ-learning protocol achieving high accuracy in predicting k_r values, which requires only a small experimental dataset (202 samples). The experimental dataset is augmented with two new datasets, with the first including 526 000 structures of the Pt(II) complexes, while the second contains 467 structures calculated using the first-principles methods, along with their corresponding simulated k_r values. Our approach is composed of two major parts: a GNN semi-supervised regression model for k_r values obtained from first-principles calculations and a supervised regression model for experimental k_r. The former model can be utilized for high throughput virtual screening (HTVS), while the latter can be used for accurate k_r predictions. This multi-level Δ-learning approach offers a way of evaluating k_r from different accuracy levels with outstanding precision. Besides, for the first time, the GNN model in this work takes into account coordination bonds and overcomes the difficulties of Pt-complex representation. Among the 526 000 generated structures, 51 new Pt(II) complexes with accurate evaluation results are presented. We expect this protocol will become a valuable tool for predicting properties of metal complexes, especially when the sample size is small, expediting the rapid development of novel OLED materials and offering guidance for the future advancement of ML models for metal complexes.