Machine Learning-Aided Screening and Design Rule Discovery for LWIR-Transparent Optical Materials

The development of low-cost, high-performance materials with enhanced transparency in the long-wavelength infrared (LWIR) region (800–1250 cm^–1/8–12.5 μm) is essential for advancing thermal imaging and sensing technologies. Traditional LWIR optics rely on costly inorganic materials, limiting their broader deployment. Here, we present a machine learning (ML)-driven framework for identifying highly LWIR-transparent hydrocarbons as candidate monomers for sulfur-organic hybrid polymers. Infrared (IR) spectra are predicted by using a communicative message-passing neural network (CMPNN), and the effect of spectral broadening (γ = 4–10) is systematically analyzed to guide flexible, use-driven γ selection. LWIR window transparency (wT) is estimated from spectra and directly predicted using regression models trained on CMPNN-derived fingerprints. These interpretable models enable the extraction of structure–property relationships, revealing chemically meaningful descriptors and substructural motifs linked to high transparency and supporting the formulation of design rules for IR-transparent materials. The approach demonstrates strong predictive accuracy and generalizability across training, test, and external data sets. Notably, several ML-prioritized motifs match experimentally validated comonomers used in IR polymer fabrication, and the model further proposes new, chemically plausible scaffolds. By bypassing the cost of density functional theory (DFT), the framework enables scalable, data-driven screening and provides chemically grounded design insights for next-generation LWIR-transparent materials.