Current failure is for legacy (1c) emulator, needs a key fix. There also appears to be an issue with the initial guess, our perpetual challenge:

The offending issue is here:

https://github.com/pgbrodrick/isofit/blob/29b0a6ddb125ed0b25784f305558720c37a16945/isofit/inversion/inverse_simple.py#L115-L122

This product clearly won't work in 4c radiance mode, only 4c transmittance.

Out of the box, simulations on the 6c emulator were taking an extremely long time:

INFO:2025-06-26,17:44:45 || common.py:__call__() |  20.02% simulations complete (elapsed: 0:28:10.708086, rate: 0:00:00.457443, eta: 1:52:42.832344)

Working this morning to confirm what the root cause is. Need to check if it's in the code or something on my end.

Hmm, how big is your LUT? I'm seeing numbers that look more like this:

INFO:2025-06-19,14:54:31 || common.py:call() | 20.06% simulations complete (elapsed: 0:00:17.743084, rate: 0:00:00.012802, eta: 0:01:10.972336)

For a LUT that looks like this:

"lut_grid": {
                "AOT550": [
                    0.001,
                    0.1009,
                    0.2008,
                    0.3007,
                    0.4006,
                    0.5005,
                    0.6004,
                    0.7003,
                    0.8002,
                    0.9001,
                    1.0
                ],
                "H2OSTR": [
                    0.4334,
                    0.6766,
                    0.9197
                ],
                "observer_zenith": [
                    4.5981,
                    8.2908
                ],
                "relative_azimuth": [
                    37.6767,
                    90.2157,
                    142.7548
                ],
                "surface_elevation_km": [
                    0.4898,
                    0.7043,
                    0.9189,
                    1.1334,
                    1.348,
                    1.5625,
                    1.7771
                ]
                }

There's definitely room for optimization (e.g. parallelize deployment over different keys), and I do expect this PR to take longer than before no matter what (we're building close to an order of magnitude more outputs), but the above looks like it's way too slow....almost like it's before I included the updates to utilize the cached H matrix.

I re-ran and run times were much faster:

INFO:2025-06-30,10:58:45 || common.py:__call__() |  90.00% simulations complete (elapsed: 0:07:47.878095, rate: 0:00:00.106336, eta: 0:00:51.986455)

Not exactly sure what the cause was. Profiling the LUT generation for a small LUT grid (size as a product of len(lut grid element): 2 * 2 * 2* 2 * 2) also confirms that things are probably running as expected. Bulk of time is the 6s simulation. The unlabeled chunk on the right are numpy array IO operations for the weight/bias matrices.

Current emulators FOR TESTING ONLY - NOT OFFICIAL: https://popo.jpl.nasa.gov/pub/PBrodrick/isofit/experimental/joint_dataset_wpoints/

Thank you Phil! FYI, using *26.npz I'm getting the error:
KeyError: 'weights_rhoatm is not a file in the archive'
stacktrace.txt

Thank you Phil! FYI, using *26.npz I'm getting the error: KeyError: 'weights_rhoatm is not a file in the archive' stacktrace.txt

Did you pull the file fresh today? It was reposted in-place (with the same name). That error will occur with the previous packed version of the .npz.

This is ready to go into dev after extensive testing and review. Thanks, @pgbrodrick! Merging now.