Still to do:

• Add tests for non-default weighting options (weighting='sunlit', weighting='insolation')
• Some basic documentation
• Make sure the problem in Can't independently specify insolation and solar zenith angle #208 has actually been fixed.

Realized that to fully support the different averaging options for coszen, we will need to modify climlab-rrtmg. The reason is that RRTMG_SW only supports scalar values for its equivalent of the irradiance factor (called adjes in RRTMG_SW code).

climlab-rrtmg v0.3 with support for non-uniform irradiance factors has now been released on conda-forge.

972721e pins the build environment to climlab-rrtmg v0.3 until #240 is solved.

Now there are a lot of tests failing for a new reason that seems to be connected to the implicit diffusion solver. I think this is a numpy v2 issue.

Lots of tests fail at one particular point: line 428 of climlab.dynamics.adv_diff_numerics which is the call to numpy.linalg.solve.

The error looks like variations on this:

ValueError: solve: Input operand 1 has a mismatch in its core dimension 0, with gufunc signature (m,m),(m,n)->(m,n) (size 1 is different from 90)

Something must have changed in the broadcasting rules for numpy.linalg.solve.

On this doc page I found

Changed in version 2.0: The b array is only treated as a shape (M,) column vector if it is exactly 1-dimensional. In all other instances it is treated as a stack of (M, K) matrices. Previously b would be treated as a stack of (M,) vectors if b.ndim was equal to a.ndim - 1.

One test climlab.tests.test_cam3rad.test_rce_radiative_forcing was failing mysteriously with an apparent energy imbalance. Some investigation revealed that a single-column RCE model with CAM3 radiation and Manabe water vapor was falling into a numerical oscillation of the temperature and moisture that led to a few W/m2 differences in the OLR at each timestep.

I haven't been able to understand why this oscillation suddenly appears in this PR, when this PR doesn't touch any of the relevant code. But simply cutting the timestep in half from 24 hours to 12 hours in that set of tests makes the problem go away, so that's what we're doing now.

Down to one single failing test about coszen stuff.