Oceans are hugely complex: modelling marine microbes is key to climate forecasts | Hacker News Reader

Oceans are hugely complex: modelling marine microbes is key to climate forecasts

nature.com

63 pointsrntn16 comments

Good article. One of the chapters of my dissertation focused on a small portion of this – I used remotely-sensed measurements of ocean surface temperature to drive mechanistic models of phytoplankton physiology to get a bit more detail on the effect that marine "heatwaves" have on their growth rates. I was able to validate this approach using another remotely-sensed dataset of ocean chlorophyll, which we'd expect to be linked to phytoplankton growth rates.

Besides being some interesting science, it was also an amazing testament to open data and high-quality scientific publishing. I was able to pull all of the ocean data from government sources for free, and leverage physiological model parameters published in open-access journals. And I was able to run the whole thing using Python/Dask on a VM in Google Cloud.

One thing that modelling seems to be missing are the lag times and genetics waiting to flourish that can allow an uptick in a higher concentration of that variation that can better adapt-take advantage of a new/changing environment, e.g. increased CO2 in atmosphere will lead to more greenery, more hardy plant life, even if it takes X years for those new plants to start to, potentially on an exponential path, take advantage of the higher CO2 and capture it.

That's actually only partially true; there's quite a bit of work that attempts to capture this idea, particularly in taxa that have relatively well-understood adaptive dynamics. Here's a good general example: https://royalsocietypublishing.org/doi/full/10.1098/rspb.201...

You are correct that much of these dynamics aren't captured in long-time-scale climate models, though. But it's being modeled quite frequently in different contexts, especially ecological ones.

And then there's the unknown of what genetics might show up down the line once increased CO2 being present starts to be selected for; including I doubt taking into account externalized factors that may occur in larger steps, like once there's more plant life to capture CO2, then animals will become larger with more food (plant and other animals), which then captures-stores more CO2, etc.

Is it even remotely possible to accurately extrapolate to what the end points all may look like? It's highly likely it's not unless you're monitoring "everything" for 100s or 1000s of years to see just how far cycles will propagate.

> then animals will become larger with more food

But see the "temperature size rule".

There's too much uncertainty in most aspects of a climate forecast that far out for it to be a meaningful exercise. The same way long term climate average is a more accurate forecasts than attempting to use a weather model two weeks out.

As with many real-world modeling exercises, it's perhaps a fool's errand to attempt to capture the entire space of possibility, unless the pure notion of capturing the domain is what's interesting.

In reality there are a number of practical constraints that place limits on the search space for these kinds of models, including ones that are likely genuinely decoupled from CO2 concentrations (think of things like mutation rates).

For a sci-fi short story I was thinking how a briefcase size solution to climate change could look like. In the story humanity would buy the "solution" to climate change from an alien high-tech travelling salesman.

What I ended up is that it could be a small vial of some bio-engineered marine microorganism designed to bloom on every ocean surface, capture sunlight and use and use the energy to capture CO2 in some chemically stable molecule and then sink to the depths. And most of the briefcase would be actually filled with a manual describing how humanity can modulate the intensity of capture to the right level with bio-engineered viruses which predate on the artificial microorganism.

Let the chemically stable molecule be teflon, and you've got yourself a sequel.

Can't unfortunately. Where would the microbe get the massive amount of fluoride required?

Likewise, where does the oxygen go? Oh well.

>capture sunlight and use and use the energy to capture CO2 in some chemically stable molecule

If only there were such a molecule!

Maybe with a couple billions of dollars in investments, scientists could find a way to make this molecule taste sweet, too.

Modelling is certainly going to be a bit cheaper than collecting comprehensive datasets from around the planet's ocean zones throughout the upper ~500m where most biological activity takes place, but without real data inputs, models tend to diverge from reality. See for example the Hawaiian Ocean Time Series project (running for ~25 years now, but just at one location, and still quite expensive):

https://en.wikipedia.org/wiki/Hawaii_Ocean_Time-series

> "The 25 year record of ocean carbon measurements at Station ALOHA document that the partial pressure of CO2 (pCO2) in the mixed layer is increasing at a rate slightly greater than the trend observed in the atmosphere. This has been accompanied by progressive decreases in seawater pH."

Thus, scientific funding needs to be split at least equally between real-world data collection and computer modelling if we want to understand what's going on. The denialist crowd would rather not know, of course, which is the equivalent of trying to avoid getting cancer by never going to see a doctor.

There was an In Our Time about plankton recently, where one of the guests talked about the Continuous Plankton Recorder, a tool using silk to trap plankton for research that was created in the 1920s and has been in continuous use since then. The housing for the silk is dragged behind boats in various places, taking cross sections of the plankton along a route.

https://hakaimagazine.com/news/the-antique-technology-still-...

https://www.bbc.co.uk/programmes/m001r1t5

When will climatologists give up and just go all-in on pure DL. Their current modelling approach is a dead end and correctly drives a lot of climate(ology) skepticism.

The current approach is to try and simulate everything from physical first principles. They started in the 70s with just the atmosphere, later added the oceans. The problem they have is that despite many claims to the contrary, their models have historically not worked and not generated accurate predictions. There are numerous examples of this like the famous "running hot" problem [0], or that since the millenium west Antarctica has cooled rather than warmed, and that no models predicted this [1]. Lack of detail around clouds are sometimes blamed. Climatologists responded by adding ever more detail to the models, in the hope that with bigger programs the predictions will come good and the (enormous) CIs will shrink.

That unfortunately didn't happen. Instead as models became more complex they diverged from each other. We are told the science is settled and only the edges need to be tied up, but in reality the science is getting less settled over time as the simulations get supposedly more accurate. It's not possible for all of them to be right but the literature (and IPCC) often just average all the models together, as if they are all equally reliable.

Now we are told that figuring out clouds (still very far from done) isn't enough, and marine microbes are also critical. Where does this end? We can't simulate the full complexity of Earth. And if more detail = less confidence continues to hold, adding this will just make the models even less useful.

On top of all that, the models are full of numerical stability bugs. NASA Model E is nearly half a million lines of FORTRAN. If you tell it to simulate a stable pre-industrial climate many runs end up catastrophically diverging to a snowball or a climate hotter than Venus [2]. Very often these bugs are papered over with hacks rather than root caused and fixed. For example, it has a comment saying "this routine makes sure that the temperature remains within reasonable bounds during the initialization process (sometimes the computed temperature iterated out in left field someplace, way outside any reasonable range)". Then they do the same for wind speeds, just using min(max(...)) to constrain that part of the simulation when it enters a feedback loop and would otherwise crater the sim. There are dozens of other serious errors that are patched up at the end with hacks. And then the model has hundreds of tunable free parameters that make it easy to overfit, some of which are explicitly computed backwards like humidity thresholds to ensure that at least some runs remain stable, instead of being set to their actual physical values.

The field needs to reset its approach. The AI field has dropped attempting to model intelligence from first principles because it didn't work, and embraced DL methods that can model any function given enough data. Google has been training and using neural nets to do short range weather forecasts. Neural climate models may not yield more accurate results (who says there's enough data) but at least there's a base of relatively solid tooling and methodologies for training, detecting overfitting, validation etc and the implementations are written by pro-level developers.

[0] https://www.nature.com/articles/d41586-022-01192-2?error=coo...

[1] https://journals.ametsoc.org/view/journals/bams/104/6/BAMS-D...

[2] https://www.netzerowatch.com/content/uploads/2023/07/Eschenb...

So, we'll save the planet by forecasting its doom by building hugely complex models which run on massive super computers which use insane quantities of energy.

I'm sure I want this to happen, but it's on the level of "100 private jets fly to a climate conference" in terms of optics.

Is this way or never. Choose one.

Some people still wants to believe that we will be able to solve some of the worst problems of the human history without spending too much resources and that it will be cheap. I don't think so.

Reply on news.ycombinator.com