The black hole of climate parameterizations

The theoretical band-aids used to "fix up" climate models after the butchering of the full theory produce a final mishmash, chunks of climate theory rounded out and connected by uncontrolled simplifications of physics too hard to solve, with details filled in with selected past behavior.

The patching up of conservation law violations and filling in for unsolvable turbulence and water dynamics are accomplished by "climate parameterizations," the rather embarrassing ad hoc-ery inherent in all present-day state-of-the-art climate modeling. Climate parameterizations are an unavoidable corollary of the GCM method. Each climate model "cell" produces a wrong answer, with no estimate of error. Somehow the sum total of these is supposed to produce a globally "right" answer. Regional climate models have even larger problems, because the climate specification on the boundaries of each "cell" is undefined to start with. If anyone comes up with the right answer using this procedure, it's strictly by luck.

These parameterizations "force closure" of the climate dynamics within each cell. "Subgrid" processes are not tracked. Instead climate parameterizations replace the missing dynamics on scales smaller than the grid resolution. They substitute for the dynamics on all but the largest spatial scales (hundreds of miles, too large to resolve even a hurricane), hiding essentially all of the chaotic behavior and most of the full hydrologic cycle (GCMs include simplified evaporation, but not condensation or precipitation dynamics). Some of these parameterizations are based on "reasonable" theoretical conjectures. But most are based on observed climate data - that is, on past climate behavior.

Briefly, there are at least three things wrong with these parameterizations.

1. Causality is eliminated within the grid cells. Cause-and-effect relationships unfolding in space and time are replaced by static, algebraic relationships. Subgrid dynamics disappears: chaotic turbulence, much of convection, condensation, cloud formation, and precipitation. Most of the self-organizing phenomena characteristic of our atmosphere are not dynamically simulated.

2. Past performance is no guarantee of future results. The climate system changes on time scales longer than modern climate data can capture. And it's also chaotic, shot through with unique, one-off events that never repeat. By using past climate data, climate parameterizations take an uncontrolled slice through the space of all possible weathers, essentially assuming that all possible weathers are represented by the time- and space-limited pool of available measurements. But this pool is restricted in time, in the spatial and temporal resolution and comprehensiveness of available data and, by its nature, cannot capture climate chaos.

Such an approach amounts to Fourier analyzing the complete climate evolution in space and time, then chopping out all but a limited range of time and spatial scales. The rest is missing, and that pesky chaos at zero frequency has been excised away. But climate processes at different spatiotemporal scales interact with one another, transferring energy, momentum, air, and water from larger scales to smaller and back, as weather features self-organize and dissipate.

3. Circularity of reasoning. To make predictions for a dynamical system, one ideally starts with a complete, defined theory, adds initial and boundary conditions, then solves for the answer. The results can be "cleanly" compared with measurements to see if the theory and any approximations made in solving it were right.

By using past climate data in defining the theory itself, we're "contaminating" the predictions of theory with the "already known answer" - cheating, in effect, although the cheater has copied a probably wrong answer. It's not a "clean" test by any means. In practice, global and regional climate models are continually adjusted to match observed climate. The resulting model looks "right," but that's an illusion. It's actually a massive case of what statisticians call "confirmation bias." The model has been adjusted to retroactively reproduce past behavior. There's no way to know if it can predict future behavior. More likely, the model will have to be readjusted again, the day after tomorrow, to "retrodict" tomorrow's weather.

The illusion of an answer. You might wonder why climate modelers ever got into what looks like a dead end. The answer is that there aren't, at present, good alternatives to this program of climate modeling approximations. Basic questions would need to be revisited and re-examined from scratch. This is a great open and urgent question in climate theory. The resources that such questions should get are instead used up in chasing illusory improvements in ever-larger and dubious GCMs. More computer power and memory can't solve this problem. It's the modeling procedure itself that's wrong. Better computers will just produce meaningless results more quickly.

While there are climate modelers and scientists guilty of overselling and misrepresenting the reliability and completeness of the GCM program, that sin pales into comparison to the main force behind this drive round and round the climate modeling cul-de-sac: it's political, not scientific. Certain political figures (not just elected politicians, but science policy and bureaucratic types as well, and the eco-fanatics) have a strong (but probably wrong) preconception of what's going on with climate. They want "correct" answers. In a larger sense, the general demand for definitive climate predictions of any kind is the more basic culprit.

The modern GCM approach to climate modeling began in the early 80s and has never left its infancy. By the early 90s, it was very prematurely "drafted" into providing pseudo-definitive climate answers. But in their current form, GCMs can never produce the answers sought or falsely claimed.

POSTSCRIPT: Essex and McKitrick discuss the full range of climate modeling fallacies in considerable, but not overly technical, detail. Leroux and Comby discuss the topic even more extensively, at greater technical depth.