Self-Diagnosing PPE Method Cuts CAM6 Ensemble Spread While Flagging Structural Error in Warm Rain Microphysics

The method decomposes the 34-parameter space into linked low-dimensional subproblems, builds conservative emulators that avoid overfitting, and explicitly tests for structural error before tightening tolerances. Controlled experiments showed that overly wide emulator uncertainty discards useful observations while tolerance of structural error shifts parameters toward compensation rather than physical fidelity. This directly addresses a recurring failure mode in Earth system model tuning where compensating biases mask model inadequacy and degrade out-of-sample projections.

By integrating constraints across subproblems the approach reconstructs a jointly plausible region instead of a single best-fit point, preserving interpretability for diagnosing which variables drive inconsistency. The 100-member PPE size remains modest, yet the iterative exclusion step demonstrably shrinks ensemble spread and improves climatological match without inflating mismatch tolerances.

Wider adoption could strengthen the traceability of parameter choices feeding into CMIP-style ensembles and national climate assessments, reducing the risk that policy-relevant quantities inherit hidden compensating errors. Next steps include testing the workflow on higher-resolution configurations and coupling it with observational uncertainty quantification pipelines already used in reanalysis production.