This is fantastic - would be really exciting to see this level of resolution making its way to global operational weather forecasts (currently at ~10km).
It's far too expensive with dubious impacts on forecast quality. Adaptive mesh approaches are far more suitable for high res global weather modeling... Why simulate the area under boring, dynamically unimportant areas?
Which is *exactly* why ultra-high resolution global weather simulation has dubious prospects for improving forecasts. When you're at spatial scales where you need to parameterize convection, there's an inherent "smoothness" to model solutions that suppresses noisy errors. If you go to cloud-resolving scales - which is needed for simulations like the ones here - you don't get the benefit of that smoothness anymore, because you need to actually resolve scales of motion that are incredibly fine. It's a losing proposition; you'll never get it "perfect", so you're much more likely to spin up an error cascade with significant impacts on forecast down the line, through things like the structure of organized convection.
But dynamically uninteresting, quasi-balanced setups and modes? There's far less to worry about in terms of the butterfly effect, and any errors you might worry about will be dwarfed by the fact that we don't have good data to assimilate in places like the remote oceans anyways.
It's also worth pointing out that the mathematics and understanding of error / perturbation growth in the atmosphere are well-understood. In fact, this fundamentally underpins how we've developed data assimilation approaches over the past two or three decades that allow us to effectively leverage new datasets such as satellite data to increase forecast quality and reliability at longer lead times. So it's somewhat trivial to actually directly quantify these "butterflies."
If we're ever going to get to the femtometer resolution required for very precise 100 day weather forecasting, we have to start somewhere, so let them waste their time. It's not as though this is part of a growing trend to abandon conventional weather and climate modeling.
Why do you think that we need "femtometer resolution" for "very precise" 100 day weather forecasting? What even is "very precise" 100 day weather forecasting? I think it's very amusing to do the math on how much memory would be required to run a crude primitive equation dycore over even the tiniest of domains at femtometer resolution :)
> It's not as though this is part of a growing trend to abandon conventional weather and climate modeling.
The thing is, there *absolutely is* a trend towards private investment in weather modeling going towards faux-moonshot ideas like cubesat constellations without demonstrated ROI and that would require evolutionary leaps forward in data assimilation, or for deep learning to replace weather models. A miniature version of this already played out with precipitation nowcasting - probably the easiest weather forecasting problem that you could approach with an AI system, yet the approaches that have been developed so far barely improve over optical flow or other simple approaches, let alone advance our capability to forecast, say, convective initiation.
The future of weather forecasting is larger ensembles (O(100-500) ensemble members, across 2-5 different models) of near-convective-resolving global models at meso-gamma (2-10 km resolution) fed into slightly more sophisticated statistical post-processing systems - almost certainly trained using simple AI/ML techniques on large-scale reforecasts of these parent model systems, or brute-forcing purely Bayesian statistical approaches.
> Why do you think that we need "femtometer resolution" for "very precise" 100 day weather forecasting?
Due to sensitive dependence on initial conditons. Even using measurements at meter resolution will cause the accuracy of a forecast to begin to break down after only a few days.
> What even is "very precise" 100 day weather forecasting?
Anywhere from accurate to exact.
> I think it's very amusing to do the math on how much memory would be required to run a crude primitive equation dycore over even the tiniest of domains at femtometer resolution
And Bill Gates thought 64K should be enough for anybody. Do you really think computers will only have a few GB of memory 50 years from now?
> there absolutely is a trend towards private investment in weather modeling going towards faux-moonshot ideas like cubesat constellations without demonstrated ROI and that would require evolutionary leaps forward in data assimilation, or for deep learning to replace weather models
This straw man does not exactly demonstrate that conventional weather and climate modeling is being abandoned anytime soon. If the unconventional private investments aren't profitable, the market will deal with them.
> The future of weather forecasting is
much like the local weather, impossible to predict with any accuracy years into the future, and yet the tools used to measure it are consistently getting more accurate, cheaper and smaller. Maybe like bottle-openers, weather sensors may superfluously start appear on everything. The more widespread the measurements, the more data descibing initial conditions, the better the forecast will be at any interval.
No, but I wouldn't mind weather measuments every cubic femtometer of the lower atmosphere and a fast enough computer with enough memory to cruch the data and accurately report what the weather will be like on 29 February.
there's a very good physical argument that this is impossible. if you want to store 1 bit per femptometer simulated, at current computer sizes, we are taking about a computer billions the size of the earth. even if you use 1 atom per bit, your computer will be almost as big as the earth. such a computer will collapse under it's own gravity.
This. No, not at all at current computer sizes, but at future computer sizes. This is the same mistake someone in the 1970's might make about billions having a smartphone today (supercomputer by their standards). Consider how everything at current computer sizes is effectively two dimensional, even stacked processors are still fundamentally 2D designs. There is still a lot of computing advancement ahead. 40 years from now they'll look back and think the same things we think when we look back 40 years, that the machines were so primitive, hardly anything could be done with them, and some will be nostalgic for them, talk about their strengths, while others will shake their heads and think even messing with the fastest workstation today is a waste of time. Just because we can't conceive of how, doesn't mean it's not possible, some day.
Unless you know it to be physically or logically impossible, you could not really know how likely it is or isn't. Ask anyone in the mid-1970's how likely it is that billions of people would be walking around with a supercomputer in their pockets, and they'd come up with all sorts of reasons why it was extremely unlikely, such as no individual would ever need so much computing power. The practicality of the precision only depends on the ability to measure and the ability to manipulate and simulate large amounts of data, both of which are extremely likely to get better, and better faster and faster, as time and technology progresses.