Grell-Devenyi scheme

Grell-Devenyi takes a different approach to convective parameterisation from Kain-Fritsch and its relatives. Rather than choosing one closure assumption and one trigger, it runs a large ensemble of combinations — on the order of a hundred or more variants spanning different closures, trigger perturbations, and entrainment rates — and returns a statistic of their tendencies, typically the mean.

The reasoning behind it is honest about the state of the field. There is no consensus on which closure is correct, and different closures perform better in different regimes. Committing to one bakes an unverifiable assumption into every timestep. Averaging over many makes the scheme’s answer less dependent on that choice.

What the ensemble is over

The variants differ mainly in the closure — what determines the convective mass flux:

• CAPE removal, as Kain-Fritsch uses: convection consumes available CAPE over an adjustment time.
• Moisture convergence: mass flux scales with the large-scale supply of moisture into the column.
• Low-level vertical velocity: mass flux tied to resolved ascent at cloud base.
• Quasi-equilibrium: convection balances the large-scale destabilisation rate.

These are genuinely different physical hypotheses about what limits convection, not variations on one. Trigger thresholds and entrainment rates are perturbed within each.

What it buys, and what it costs

The benefit is reduced sensitivity to a single arbitrary choice, and a scheme that behaves more consistently across regimes than any of its members would alone. It is a natural choice for sensitivity studies, where the question is whether a regional precipitation signal survives convective parameterisation uncertainty rather than what one scheme predicts.

The costs are worth stating plainly. It is more expensive than a single-closure scheme, though less than the member count suggests since the members share most of their computation. More importantly, the ensemble mean is not itself a physically consistent solution. Averaging the tendencies of a hundred plausible convective states gives a state none of them would have produced — which tends to smooth out extremes and can weaken the very events a study is about. The ensemble spread is arguably the more informative output, and it is usually discarded.

There is also a subtler issue: the spread reflects the range of formulations someone chose to include, not a calibrated uncertainty. Treating it as a probability distribution over convective behaviour would be reading more into it than it supports.

See also: cumulus convection schemes for the full comparison, Grell-Freitas for the scale-aware successor, Grell 3D, and Kain-Fritsch for the single-closure alternative.