Dec 2022: "Official 2.0 release" corresponding to submitted GMD (previously a JAMES preprint earlier in the year) article describing RRTMGP-NN implementation in ecRad and prognostic testing in the IFS. I make no claims on code maturity/usability but some useful new methods for training more accurate gas optics neural networks are demonstrated in /examples/rrtmgp-nn-training:
- Monitor flux errors with respect to LBL solution (RFMIP data) while training, by calling radiation scheme with models as they are being trained, and early stop on those
- hybrid loss function to minimize the error in the difference in outputs y associated with different perturbation experiments (proxy for radiative forcing errors)