ILAMB cmip7 baselines blocked by ZeroDivisionError in ilamb3 spatial_distribution (Taylor score)

[lewisjared]

Summary

The cmip7 regression baselines for gpp-wecann and gpp-fluxnet2015 cannot be minted: the ILAMB run fails during reference-intermediate generation with a ZeroDivisionError in ilamb3's spatial-distribution (Taylor) analysis. Their cmip6 baselines mint fine; only the CMIP7 path is affected.

Where

ilamb3/analysis/spatial_distribution.py:133:

taylor_score = 4 * (1 + corr) / ((norm_std + 1 / norm_std) ** 2 * 2)

norm_std is the model field's normalised spatial standard deviation. On the CMIP7-converted CanESM5 data for these gpp diagnostics it evaluates to 0, so 1 / norm_std raises ZeroDivisionError. A zero/degenerate spatial standard deviation should be handled (e.g. guard the reciprocal and emit NaN) rather than crashing the whole analysis.

Reproduce

ref test-cases fetch --diagnostic gpp-fluxnet2015
ref test-cases mint --provider ilamb --diagnostic gpp-fluxnet2015
# -> ilamb/gpp-fluxnet2015/cmip7: source stage failed during mint:
#    Reference intermediate data was not generated.  (ZeroDivisionError)

Traceback tail:

File ".../ilamb3/run.py", line 656, in run_analyses
    df, ds_ref, ds_com = a(ref, com)
File ".../ilamb3/analysis/spatial_distribution.py", line 133, in __call__
    taylor_score = 4 * (1 + corr) / ((norm_std + 1 / norm_std) ** 2 * 2)
ZeroDivisionError: float division by zero

Impact

gpp-wecann/cmip7 and gpp-fluxnet2015/cmip7 ship with cmip6 baselines only until this is resolved upstream in ilamb3. Likely affects any diagnostic whose CMIP7 model field is spatially constant over the comparison region.

[lewisjared]

Root cause found and the primary defect is fixed in PR #779 (commit 0f9ef5f).

The CMIP6->CMIP7 converter applied least_significant_digit=3 (~1e-3 absolute precision) uniformly to every variable. gpp magnitudes (~1e-8 kg m-2 s-1) are five orders of magnitude below that floor, so the entire converted gpp field rounded to zero. A spatially uniform (zero-variance) field then made the Taylor spatial-distribution score divide by 1/norm_std == 0. This is why only the two gpp diagnostics failed, only on the CMIP7 path, and only after conversion (the native CMIP6 gpp is untouched). The fix skips decimal-place quantization for gpp; other variables convert byte-identically. CMIP7 baselines for gpp-wecann and gpp-fluxnet2015 now mint.

Remaining scope for this issue (defense-in-depth, not blocking): ilamb3's spatial_distribution should guard norm_std == 0 rather than raising ZeroDivisionError, so a genuinely degenerate field degrades gracefully. Also worth noting: other very-small-magnitude flux variables (e.g. nbp ~1e-9, fgco2) would hit the same converter zeroing if they were ever run through a spatial analysis; they pass today only because their diagnostics use global-mean references. A general per-variable relative-precision encoding would be the durable converter fix.