Machine Learning-Assisted Synthesis-to-Optics Screening of Ag@SiO2/Polymer Nanocomposites for Visible Spectrum Negative Effective Permittivity

Machine learning (ML)-assisted design of epsilon-negative polymer nanocomposites requires a clear connection between experimentally controllable synthesis parameters, core–shell nanoparticle geometry, and the resulting effective optical response. The targeted optical response is unusual because the polymer film is predicted to exhibit near-zero or negative real effective permittivity in selected visible spectrum regions, arising from Ag core plasmonic polarizability, SiO2-mediated dielectric spacing, nanoparticle filling factor, and effective medium coupling rather than from the intrinsic polymer matrix. In this study, a two-stage ML-assisted synthesis-to-optics framework is developed for Ag@SiO2 core–shell nanoparticle/polymer composite films intended for visible spectrum effective permittivity screening. In the first stage, Stöber synthesis parameters, including water volume, ethanol volume, TEOS content, catalyst volume, reaction time, Ag nanoparticle size, and Ag nanoparticle concentration, were used to predict SiO2 shell thickness. In the second stage, Ag core size, SiO2 shell thickness, wavelength, and nanoparticle filling factor were used to screen the real effective permittivity of Ag@SiO2/polymer nanocomposites within an effective medium design space. Using a duplicate-aware validation workflow, Gradient Boosting provided the strongest held-out test performance for shell thickness prediction, with a test R2 of 0.8997, MAE of 7.1822 nm, RMSE of 8.8344 nm, and cross-validation R2 of 0.5371 ± 0.4648. The relatively large cross-validation variability indicates that the model is useful for interpolation-based synthesis screening but should not be interpreted as fully robust across heterogeneous literature-derived data. For the optical response task, the highest held-out test performance was obtained by a Decision Tree model (test R2 = 0.7586), but cross-validation results were unstable, indicating that the epsilon model should be interpreted as a design space screening tool rather than a generalizable predictor. Design window analysis identified candidate negative effective permittivity regions primarily at 400 nm and high nanoparticle filling factor, with predicted Re(εeff) values ranging from −5.4229 to −0.2086 across selected windows. The main contribution of this work is the treatment of SiO2 shell thickness as a bridge variable between Stöber-derived synthesis control and effective permittivity screening. Experimental validation remains necessary to confirm the predicted design windows, particularly because shell uniformity, Ag core polydispersity, nanoparticle aggregation, polymer dispersion, high-filling-factor feasibility, and effective medium validity can strongly influence the measured optical response.