Smart molecular design of NIR-II organic fluorophores through self-driven iterative evolution, deep learning, and fragment-based assembly.

Near-infrared II (NIR-II, 1000-1700 nm) emissive molecules are highly valued for biomedical imaging, phototherapy, and optoelectronic applications due to their deep tissue penetration, reduced autofluorescence, and high signal-to-noise ratio. However, their rational design remains challenging, as conventional discovery relies heavily on labor-intensive synthesis, limited quantum chemical calculations, and inefficient trial-and-error exploration. To overcome these limitations, an iterative AI-driven molecular evolution strategy (AI4NIR-II 1.0) is introduced that integrates time-dependent density functional theory (TDDFT), transformer-based predictive modeling, and generative molecular design. Starting from donor-acceptor-donor (D-A-D) and donor-donor-acceptor-donor-donor (D-D-A-D-D) fragment scaffolds, a predictive model for both absorption and emission properties was trained based on a high quality dataset containing ∼16,000 molecules annotated by the optimally-tuned range-separated LC-ωHPBE* (OTRS) functional. A fine-tuned generative model was subsequently incorporated into a self-refining-loop workflow that cycles through molecular generation, property screening, and dataset augmentation. The model achieves excellent predictive performance for both emission and absorption properties, that is mean absolute errors of 19 nm for emission peak wavelengths and of 11 nm for absorption peak wavelengths, and correlation coefficient (