ESRL is terrible. I know of much, much better models (forget names). They’re described in peer-reviewed journals (don’t recall cites, paywalled so haven’t read) but that data (proprietary format, floppy disk by request) shows how bad the ESRL fall 2016 validation really is (can’t really give specifics, no time).
Thanks for sharing.
Actually ESRL did not even begin their archive until nine months later, on July 27th of 2017. ESRL did not go ‘whole Arctic’ until the 22nd of August 2017. The web site was first opened on August 23rd. ESRL products, as labelled in 20 pt font on all 37 daily graphics, are experimental. And that product mix is still being revised daily into September.
Only as the upcoming refreezing season progresses will sufficient product accumulate for evaluation, after which the much-published PI will publish another journal publication. For now, let’s not confuse validation with placeholder (lorem ipsum dolor...)
In my view, there is only do, or not do.
In physics, if no experimentally testable prediction is coming out the door, you’re not doing. For climate science, forecasting means an .nc data file and its graphic stably posted without delay or registration barriers to the open internet. It does not mean twitter media or a dead link to little endian Word compilation in a non-peer reviewed supplement of an outdated paywalled journal article. At this time, only ESRL and Hycom are predicting the Arctic Ocean ice; only ESRL provides the data files.
ESRL is about forecasting out to D5 (and for some observables, out to D10); Hycom goes to D7. At this time, both offer archives but neither posts hindcasts or reanalysis. To make a good forecast, ESRL needs a decent D0, the initial state. So there are two separate validation issues: how good is the D0 and, given a spot-on D0, how well does their coupled air-ice-ocean model move it forward to D5? However the nature of the Arctic ice is such that neither is easily tested:
We’ve agreed to agree (with some holdouts) that nobody has the slightest idea what the ice thickness really is, within 50% error, across the Arctic Ocean on a given day. We’re not on solid ground when 2m of observed “buttery ice” at the North Pole on 05 Aug 17 is held equivalent to 2m brittle ice in the Lincoln Sea (show me the Arctic-wide buttery ice map). There is a week of accurate but limited 2017 swath data off Alert but so far it’s unpublished.
In 2019, the Polarstern will spend a winter collecting real ice thicknesses by helicopters 35 km to the sides of its drift. That data may be an uncomfortable fit to certain model products followed on the forums (the ones that make testable predictions, not all do).
https://www.theatlantic.com/magazine/archive/2017/10/a-year-on-ice/537912/Bottom melt? Nobody is down there in scuba gear; can predictions be assessed from a few broken-down buoys and moorings in wrong places? Rain gauges? Not a single one for an area the size of Europe. We go by droplets and soggy snow on a couple of web cams. That's a budget issue though, not string theory.
So while ESRL has gone all-in on ice thickness prediction, the best part really is their release of all the contributing components to that calculation (e.g. ice-to-ocean thermal flux daily graphic and its underlying .nc data). There’s no obligation to use all their inputs or computational pipeline. Anyone can stub in an improved ingredient if they have it, or use it for validation.
As examples, the UH AMSR2 3.125 km sea ice concentration is a far better product than some of the 25 km resolution competition. It’s backed by a careful journal article; the underlying .nc data accompanies the daily image; there’s no mickey-mouse registration barrier or ftp passwords. ESRL uses AMSR2 but not this one. So for a value-added hybrid concentration, the UH can be used as initial state until it runs out (day before) and ESRL can pick it up and take it forward to D5.
The time series shown below takes 20 archived ESRL initial states for ice thickness (which go up to today, Sept 10th) and pushes that out to the expected change in ice thickness on Sept 15th, D5. The archive is a bit awkward to use for this but a volunteer here is restructuring it in the auto-cloud.
The D5 image below has pixel counts next each palette tile. Taking the dot product of count and thickness bin to estimate volume change, it emerges that Sept 15 is a wash, relative to the 10th. There's slightly more melt than freeze but the difference is minor compared to error issues. This has been a very flat minimum and ESRL sees that continuing to mid-month.
