RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación
Artículo científico (JournalArticle)
Bridging machine learning predictions and perceived color in gold nanoclusters through spectral reconstruction and CIE colorimetry
[EN] Gold nanoclusters (AuNCs) exhibit photoluminescence strongly dependent on their atomic structure, ligand environment, and synthesis conditions, making the rational design of targeted emission characteristics particularly challenging. In this study, we have developed a machine-learning framework to predict the maximum emission wavelength (¿em) of AuNCs and subsequently reconstruct approximate emission spectra using a size-dependent full width at half maximum (FWHM) model. A curated database of 802 literature-derived AuNC samples was compiled, integrating ligand chemistry, synthesis conditions, physicochemical properties, and optical characteristics. An ensemble XGBoost model achieved an R² of 0.9228 and a mean absolute percentage error of 1.25% for thiolate-protected systems. Predictive performance was highest in the visible spectral region, where data representation was greatest, while uncertainty increased in sparsely represented regions. Predicted ¿em values were then transformed into approximate emission spectra using a size-dependent FWHM model, which were subsequently converted into Commission Internationale de l¿Éclairage (CIE) L*a*b* color coordinates. Experimental validation with independently synthesized sulfur-containing ligand-protected AuNCs confirmed the ability of the framework to predict ¿em and reproduce the dominant colorimetric characteristics of the experimental systems. The results demonstrate that AuNCs exhibiting similar ¿em may display distinct perceived colors because of differences in spectral broadening, highlighting the limitations of approaches predicting only ¿em. The integration of machine learning with mathematical spectral reconstruction establishes a direct relationship between computational predictions and human color perception, offering a powerful strategy for the data-driven design of nanomaterials for applications in bioimaging, colorimetric sensing, and advanced optical technologies., This study was supported by Generalitat Valenciana under the
Prometeo programme (CIPROM/2022/57). LFS acknowledges support
from MICIU//AEI/10.13039/501100011033 and FSE + , under her
Ramon y Cajal contract (RYC2023 043859-I). The authors also
acknowledge CRUE-Universitat Politecnica de Valencia for funding the
open access publication charges.