Interpretable Machine Learning-Assisted Multi-Parameter Optimization of Ag@SiO2 Core–Shell Nanofluids for Spectral Splitting Photovoltaic/Thermal Systems

Nanofluid spectral filters can improve full-spectrum utilization in photovoltaic/thermal (PV/T) systems, but the coupled effects of core–shell geometry and optical depth make conventional trial-and-error optimization inefficient. Here, a physics-based and interpretable machine learning framework is developed for the multi-parameter optimization of water-based Ag@SiO2 core–shell nanofluid filters. A high-throughput dataset was generated by coupling generalized Lorenz–Mie theory, multilayer Monte Carlo radiative transfer, and a PV/T performance model under prescribed cell-temperature and collector assumptions, with Ag core diameter, SiO2 shell thickness, nanofluid layer thickness, and whole-particle volume fraction as design variables. The coated-sphere implementation was verified through both its zero-shell limiting-case reduction and an independent finite-shell benchmark against MiePlot v4.6.19. Six regression algorithms were evaluated using geometry-grouped data splitting and cross-validation, and the multilayer perceptron was selected as the final surrogate. SHAP analysis attributed 31.6%, 27.2%, 25.5%, and 15.8% of the global importance of the merit function to volume fraction, layer thickness, shell thickness, and core diameter, respectively, while revealing saturation, threshold, and nonmonotonic parameter effects. Five independent genetic-algorithm runs converged to a candidate numerical design with an approximately 48 nm core diameter, 11 nm shell thickness, 18.4 mm fluid layer, and a whole-particle volume fraction of 2.2 × 10−5, yielding a surrogate-predicted merit function of 1.48014. High-fidelity recalculation yielded ηpv = 8.96691%, ηth = 43.62598%, and MF = 1.479478, while confirming an absorption-dominated particle resonance near 0.43 μm, low reflectance over most of the spectrum, and enhanced long-wavelength absorption. This framework enables efficient and physically interpretable exploration of the continuous design space for core–shell nanofluid PV/T filters.