Interpretable neural operators enable topology-optimized cooling for concentrating photovoltaics

Concentrating photovoltaic systems generate high electrical output but also severe heat loads that reduce efficiency and reliability. Here, we report an interpretable neural operator framework for designing topology-optimized cooling channels for concentrating photovoltaics. A deep operator network learns the mapping from variable-depth flow field geometries to temperature and pressure fields, enabling rapid design screening while reducing computational energy use by more than 99% compared with conventional computational fluid dynamics (CFD). The optimized topology is fabricated by high-resolution stereolithography and tested under concentrated illumination. It lowers peak surface temperature from about 60°C to 54°C at 10 suns and is supported by numerical results showing a 15.7% increase in Nusselt number. Physical interpretation shows that topological undulations induce wall-normal secondary flows and improve field synergy, actively disrupting the thermal boundary layer. This framework provides a general route for interpretable, data-driven thermal fluid design in high-heat-flux energy devices.