Glass windows are the most energy-inefficient part of
buildings,
which triggers the ongoing chasing of energy-efficient transparent
radiative cooling (TRC) metamaterials on glasses that simultaneously
maintain high visible (VIS) transparency, block near-infrared (NIR)
solar radiation, and emit thermal energy through the atmosphere window
(AW). However, the stringent multispectral regulation remains challenging
since it involves with huge parameter spaces and significant interactions
among different bands. Additionally, most TRC metamaterials require
a top ∼50 μm polymer for high emissivity in the AW, which
will reduce the VIS transparency and suffer from aging issue. Here,
we employ the deep reinforcement learning (DRL) method, leveraging
its robust material screening and structure optimization capabilities,
to design a five-layer submicrometer dielectric multilayer, composed
of two stacked materials, as polymer-free TRC metamaterial on glass
or meta-glass for short. Utilizing a plain glass substrate with high
emission in the AW, our meta-glass demonstrates an ultrahigh angular-independent
(<60°) VIS transmissivity against the state-of-the-art, i.e.,
86% (92.7% in theory), with ∼48% NIR reflectivity and ∼89%
AW emissivity in experiment. In outdoor experiments at ambient temperatures
of ∼10 and ∼20 °C, with solar irradiances reaching
around 780 and 850 W/m2, our meta-glass achieves a floor
temperature reduction of 8.9 and 12.7 °C, respectively, compared
to uncoated glass. Furthermore, we achieved customization of meta-glasses
with varying transparency levels while maintaining high NIR reflectance
by DRL. Our meta-glass exhibits an extraordinary building energy saving
potential in most climate zones. This work provides a valuable reference
for the advancement of TRC and the design of multispectral metamaterials.