Radiation Schemes
Radiation schemes compute how shortwave and longwave radiation propagate through the atmosphere and what heating rate that implies at each level. They are the model’s connection to its actual energy source, and effectively everything else responds to their output: surface temperature, boundary-layer growth, convective instability, and the circulation that follows from all three.
Solving radiative transfer exactly is far too expensive, because absorption by water vapour, CO2, ozone, and the rest varies enormously with wavelength across millions of spectral lines. Every operational scheme is an approximation to that spectral integral, and the schemes differ mostly in how they make the approximation.
Where the cost goes, and what gets sacrificed
Two decisions dominate.
Spectral treatment. The correlated-k method used by RRTMG groups wavelengths by absorption strength rather than by wavelength, reducing millions of lines to a few hundred pseudo-intervals. It is accurate and comparatively expensive, and it is the standard choice where getting the longwave right matters.
Angular treatment. Almost all schemes are two-stream: radiation travels up and down and nowhere else. This is a genuine approximation. It cannot represent horizontal transfer between columns, which means it cannot handle a cloud shading its neighbour, or illumination of a slope by a nearby sunlit valley wall. At coarse resolution the error is small. At the micro-scale, and in complex terrain, it stops being negligible — the reason PALM-4U carries a separate radiative transfer treatment for street canyons.
Because they are expensive, radiation schemes are usually called every 10-30 model minutes rather than every time step. That is normally fine, but it aliases the diurnal cycle if the interval is set carelessly, and it interacts badly with rapidly evolving cloud.
The dependency that causes most trouble
A radiation scheme is only as good as the cloud field handed to it. Cloud dominates both shortwave reflection and longwave trapping, and radiation does not compute cloud — it receives it from the microphysics and the convection scheme via convective cloud fraction.
This is why radiation, microphysics, and convection have to be configured as a set. In WRF, changing the cumulus scheme without checking what the radiation scheme assumes about cloud fraction and overlap is a reliable way to acquire a temperature bias that appears to come from nowhere. The scheme is rarely the problem; the interface between schemes usually is.
The cloud overlap assumption deserves specific attention, since it lives in the radiation code, is rarely mentioned in configuration discussions, and changes total cloud cover substantially for deep convective columns.
See also: RRTMG for the common choice, surface schemes for what receives the surface fluxes, convective cloud fraction for the handoff, and the model physics map.