Evaluating TabPFN for reconstructing spectral solar irradiance from broadband observations

High-resolution terrestrial solar radiation data is fundamental for understanding wavelength-dependent atmospheric attenuation, which governs photovoltaic efficiency, agricultural photomorphogenesis, and radiative forcing. However, the high cost and maintenance requirements of precision spectroradiometers prohibit widespread deployment. This study utilizes TabPFN—a transformer-based tabular foundation model—to reconstruct surface solar spectral irradiance at a 1-nm resolution using broadband observations. By parameterizing the model with meteorological variables, solar geometry, and atmospheric indices, we simultaneously reconstruct the spectral components of global, beam, and diffuse irradiance. Validation against three years of field observations demonstrates that TabPFN significantly outperforms conventional machine-learning benchmarks. The reconstructed spectra preserve physical integral consistency, with broadband summations over visible, near-infrared, and photosynthetically active radiation bands showing excellent agreement with observations. Furthermore, the model exhibits high robustness across clear, cloudy, and overcast skies, providing a computationally efficient solution for generating accurate spectral data for atmospheric and solar energy research.