Machine learning-accelerated design of near-infrared TADF emitters beyond conventional scaffolds

Abstract

Near-infrared thermally activated delayed fluorescence (NIR-TADF) emitters hold great potential for bioimaging, optical communication, and advanced display technologies, but their design remains challenging due to the energy gap law, which intrinsically limits efficiency at low emission energies. Here, we present a machine learning (ML)-guided inverse design strategy that accelerates the discovery of high-performance NIR-TADF molecules. A chemically diverse library of 1849 candidate structures was generated by fragment recombination of known donor–acceptor and multi-resonant motifs. A property-prediction ML model rapidly screened this library for low singlet–triplet gaps (∆EST) and long-wavelength emission, followed by density functional theory validation of shortlisted candidates. Eighteen previously unexplored molecules were identified with emission wavelengths beyond 700 nm and ∆EST < 0.2 eV, demonstrating the ability to break the conventional trade-off between color purity and radiative efficiency. Photophysical simulations confirm their potential for optoelectronic applications. This work establishes a transferable ML-assisted design framework applicable to other emissive systems, opening new directions for computationally driven molecular discovery.