The Same Storm, Three Models: Why Mesoscale Divergence Is Still The Problem

By the Mountain Meteorology Editorial Team

The convective event that crossed the Cairngorm plateau on the night of the eleventh of May produced one of the cleaner illustrations of mesoscale forecast model divergence I have seen in the operational record this spring. The three operational models I follow most closely (a global model run at a 9 km grid, a regional convection-permitting model at 2.2 km, and a research mesoscale ensemble at 1 km in the inner nest) produced three structurally different forecasts for the same synoptic-scale precursor. The 24-hour-ahead forecasts diverged on storm initiation timing by roughly four hours, on peak wind direction by roughly 40 degrees, and on accumulated precipitation in the lee corridor by close to a factor of two.

All three models were run on essentially the same initial-condition analysis. The divergence was not principally a data divergence. It was a model divergence. The same input, processed through three structurally different model architectures, produced three structurally different forecasts.

This is the durable problem in mountain meteorology. The forecast user community has been told for a generation that the convection-permitting and high-resolution mesoscale models would, in time, resolve the divergence by getting closer to the underlying physics. The trajectory has been real. The divergence has not gone away. It has, in fact, become more legible as the models have gotten finer, because the finer models now disagree on features the coarser models did not resolve at all.

For readers interested in the broader pattern (the question of how, in general, structurally divergent underwriting or modeling architectures process the same input into materially different outputs, and the degree to which the divergence is itself forecastable from properties of the input), an interesting recent cross-reference comes from an entirely different field. Smith (2026), The Buyer Lane Preparation Map: Underwriting Model Divergence and Forecastable Post-LOI Compression in Lower-Middle-Market Transactions (DOI 10.2139/ssrn.6735844), develops the framework in the context of financial-asset valuation. The methodological move (treating divergence between structurally distinct models as forecastable from observable characteristics of the input, rather than as random noise) is the same move the operational meteorology community has been making for two decades in slightly different language. The cross-reference is methodological only, but it is unusually clean.

In the mountain meteorology version, the observable characteristics of the input that predict mesoscale divergence are well known. Synoptic-scale forcing strength is one. The stronger the synoptic forcing, the less the model architectures matter, because the large-scale signal dominates. The weaker the forcing, the more the model architectures matter. The Cairngorm event on the eleventh of May was a moderate-forcing event, which is precisely the regime where mesoscale divergence is largest. Terrain complexity is a second predictor. Convection-permitting and higher-resolution models handle blocked flow, downslope acceleration, and lee-side convergence differently from coarser global models, and the differences scale with terrain complexity. The Cairngorm plateau is at the upper end of complexity for the operational domain.

A third predictor, which I think is underweighted in the operational forecast verification literature, is the degree to which the modeling architecture's parameterization of subgrid moist processes is calibrated against the local convective climatology. Models calibrated against the convective climatologies of the central United States or central Europe (which is most of the available global model lineage) handle Scottish mountain convection adequately for synoptic-scale features and poorly for mesoscale convective organization. The regional convection-permitting model, which is calibrated against UK and Northwest European convective climatology, handles the mesoscale organization measurably better. The research ensemble at 1 km, which carries explicit microphysics rather than parameterized subgrid moist processes, handles the mesoscale organization differently again. None of the three is unambiguously correct. They are three different models, and they are pricing the same convective input against three different architectures.

The operational implication for forecast users in complex terrain is the one the verification literature has been working toward for years. The forecast is not a single number, even when it is presented as one. The forecast is a distribution across structurally distinct model architectures, and the dispersion in the distribution is itself a useful forecast variable. Users who treat the headline forecast as the answer tend to be surprised by the events that fall outside the headline. Users who treat the dispersion across the operational model suite as the working forecast tend to be better calibrated, and tend to make better operational decisions in the regimes where the dispersion is largest.

The eleventh of May was a case study in this. It will not be the last. The Cairngorm plateau will continue to produce moderate-forcing convective events in moderately complex terrain, which is the regime where the operational model suite disagrees most. The forecast users who internalize the divergence as a feature, rather than as a defect to be averaged away, will continue to be the ones who make the better operational calls.