Machine-learning-assisted design of energy-saving windows with high near-infrared shielding properties

Nanocomposite films based on Cesium tungsten oxide (CWO) and Indium tin oxide (ITO) nanoparticles provide a broad space for adjusting the optical properties of energy-saving windows due to their unique near-infrared absorption properties. This property has led to great research interest in the field of energy-saving windows for such materials. The optical properties of energy-saving windows are mainly determined by localized surface plasmon resonance (LSPR) of the nanoparticles, and thus they are sensitive to the variation of the geometrical parameters of the nanoparticles. Typically, the computational cost of the design of specific optical properties and iterative optimization of the geometrical parameters is expensive and time-consuming. In this study, we combine machine learning and radiative transfer calculations to achieve targeted design energy-saving windows. By adjusting the shape, material, and geometric parameters of nanoparticles, an analysis model can be established from the geometric parameters of nanoparticles to the properties of energy-saving windows. Then, a machine learning model of bidirectional deep neural network is developed to achieve accurate prediction of optical evaluation parameters (visible transmittance (T_lum), near-infrared (NIR) transmittance (T_NIR), solar radiation transmittance (T_sol), and the Figure of Merit (FOM)) for energy-saving windows, as well as inverse design of geometric parameters of nanoparticles (CWO and ITO). The results indicate that our machine learning model achieved forward prediction of energy-saving window optical properties with an accuracy of over 99 % and inverse geometric parameter design with an accuracy of over 93 %. Overall, this work provides a broadly appropriate and computationally efficient method for evaluating and designing the properties of energy-saving windows.