Development of data-driven property prediction approaches for the discovery of thermally activated delayed fluorescence materials

Abstract

Deep learning has the potential to accelerate material discovery by rapidly and accurately identifying materials with desired properties within a vast chemical space, as it can effectively capture non-linear relationships between molecular and quantum chemical properties. In this thesis, we present a series of deep learning-based approaches aimed at efficiently discovering thermally activated delayed fluorescence (TADF) materials. First, we explore suitable molecular representations and model architectures for deep learning-based molecular property prediction tasks. Subsequently, we introduce a novel model architecture, namely dual geometric message passing, designed to provide precise predictions for quantum chemical properties that rely on two molecular geometries. Additionally, we propose a training framework capable of predicting high-level quantum properties using only easy-to-obtain molecular geometries. Building upon this framework, we introduce a novel approach for predicting barrier heights by integrating it with the dual geometric message passing architecture. Lastly, we propose a deep learning-based chemical similarity that enables the selection of molecules with similar quantum chemical properties to those found in existing TADF materials. Importantly, our approaches do not rely on domain-specific chemical properties unique to the field of TADF materials discovery. As such, we anticipate that our methods are readily applicable to various other materials discovery fields, extending well beyond the realm of TADF materials.