Stochastically Perturbed Parameter scheme
SPP perturbs selected parameters inside physics packages using a random pattern correlated in space and time. In WRF it can be applied to the Grell-Freitas convection scheme, the MYNN boundary-layer scheme, and the RUC land-surface model, enabled through spp_conv=1, spp_pbl=1, and spp_lsm=1 respectively. The perturbation fields are written to history output as pattern_spp_conv, pattern_spp_pbl, and pattern_spp_lsm. Implementation follows Jankov et al., Monthly Weather Review, 2017.
The reasoning
A parameterisation replaces unresolved physics with a formulation containing tuned constants. Those constants are uncertain — chosen from limited observations, valid for the regimes they were fitted in, and applied everywhere regardless. Running with fixed values asserts a confidence nobody has.
A deterministic run therefore understates uncertainty in a specific way. It is not that the answer is noisy; it is that a plausible alternative parameter set would have given a systematically different answer, and running once hides that entirely.
SPP represents this by letting parameters vary within their plausible ranges, smoothly in space and time so the perturbations behave like a coherent alternative formulation rather than like noise. The correlation structure is what makes it meaningful — uncorrelated random perturbations would average out and change nothing.
Compared with the alternatives
The Grell-Devenyi ensemble addresses the same problem by averaging over many closures and returning the mean, which produces a single physically inconsistent state and discards the spread. SPP keeps each member internally consistent and produces spread as output. For quantifying uncertainty rather than reducing sensitivity to it, that is the better structure.
The costs are real. The perturbation ranges have to be specified, and they are themselves uncertain — there is no principled way to set them, so the resulting spread is a function of a choice rather than a calibrated uncertainty. Reading it as a probability distribution over outcomes claims more than the method supports.
And it only makes sense as an ensemble. A single perturbed run is just a run with different parameters, which is worse than useless if read as deterministic.
See also: Grell-Devenyi for the averaging approach, cumulus schemes, MYNN, and the physics map.