Climate ChangeRecent Post

RealClimate: Digital Twinge

A couple of weeks ago the EU announced that they were funding a project called DestinE (Destination Earth) to build ‘digital twins’ of the Earth System to support policy making and rapid reaction to weather and climate events.

While the term ‘digitial twin’ has a long history in the engineering world, it’s only recently been applied to Earth System Modeling, and is intended (I surmise, as does Bryan Lawrence) to denote something more than the modeling of either weather or climate that we’ve been doing for years. But what exactly? And is it an achievable goal or just a rebranding effort of things that are happening anyway?

To summarise the proposals, here are two videos that give the vision for DestinE as seen from ECMWF. The first from PI. Dr. Peter Bauer (on Youtube – not embeddable for some reason) and one by Dr. Irina Sandu:

The triplet nature of the digital twin

There appear to be three strands to the digital twin concept as conceived by the EU (see also Bauer et al (2021)). The first is an upgrade to the basic weather forecast machinery at ECMWF to higher horizontal resolution. This is perhaps an acceleration of work that would likely have been done anyway and, based on past performance, is likely to pay dividends in improved weather forecasts, and more realistic reanalyses (ERA6?) eventually. The reason we can be confident that this is practical is that we have seen dramatic increases in weather forecast accuracy over the last couple of decades in part driven by the increases in resolution (as well as improvements in process modeling and increased observational data ingestion).

Skill score improvements at ECMWF from 1980 to present for 3, 5, 7 and 10 day forecasts (note the expanding vertical scale).

The connection between higher resolution and better weather forecasts lies in the decrease in numerical diffusion, and better ability to track the large scale synoptic situation, including the fronts and atmospheric rivers etc. before the chaotic nature of the weather renders forecasts unreliable. In other words, weather forecast accuracy relies heavily on the simulation of atmospheric dynamics, and if that improves, so will the weather forecast (up to a point).

The second strand is that the new digital twins for climate will be better linked to impact models across a suite of fields – from urban planning, water resource management, coastal resources, emergency planning, infrastructure design etc. This would be undoubtedly be a good thing to have. However, this is a large increase in complexity and one that does not have a prior track record of success to build on. The issues here are that a) climate forecasts need to be accurate on the scale of the local impact, and b) that some way can be found to connect up the myriad impact models to the forecasts at scale despite the massive heterogeneity in tools, inputs, platforms, users etc. This is (to my mind) a huge challenge even if we assume that point (a) is deliverable (but I’ll return to that below).

The third strand is heavily user-focused. A step change in the quality of data visualization and user interfaces, combined with a greater degree of interactivity driven by users themselves as opposed to the developers. This is interesting, but raises many questions about what would really be doable.

En-twin-ing climate and weather

However, in the presentations above there is a large and mostly unstated assumption that revolves around the question of whether an improvement in climate forecasts, particularly at the decadal scale, would follow from the increased focus on higher resolution in climate models. Many people working in numerical weather prediction, having seen the improvements that higher resolution has brought, often claim that climate predictions will become more accurate at about the same rate. But this neglects the fact that the factors that provide predictability at either the decadal or multi-decadal timescale are not the same as the drivers of accuracy in weather forecasts. In climate models, these revolve around the climate forcings and climate feedbacks, not so much the quality of the dynamics.

This is not to say that better dynamics have no impact on climate feedbacks – indeed, there are many potential paths for this to make a difference. But the dominant uncertainty in climate feedbacks involve cloud processes (micro-physics, macro-physics, aerosol-cloud interactions) that are still many decades away from being fully captured from first principles in global models. Take the CMIP6 models that we have written about before (#NotAllModels, Sensitive but Unclassified, Part II): the biggest causes of diverging climate projections relates to Southern Ocean cloud feedbacks (Zelinka et al , 2020), which is unconnected to resolution. In fact, there is no evidence that I am aware of that suggests that increasing resolution (in the current climate model regime*) reduces uncertainty in climate sensitivity.

Additionally, there is a very basic assumption that (given the forcings) more accurate decadal/multi-decadal forecasts of important impacts are possible at higher resolution. Results with large ensembles however, show that, even for 50 year trends in spatial precipitation patterns at the continental scale, there is a huge range of variability just from the unforced, weather, component (Deser et al, 2015) – and this is unlikely to change with higher resolution models, it may even increase!

However, if we accept that while higher resolution likely won’t reduce uncertainty in global warming, it might provide better estimates of local impacts, given any particular level of global warming. But this would suggest a hybrid design, whereby SST/sea ice projections from climate models could be used as input to weather models to better assess the consequences for the weather regime. Frankly, I am surprised this isn’t done more, but it doesn’t appear to be considered here.

*Eventually, when we are modeling everything from first principles, we would expect convergence as resolution increases, but we are not yet there. Until then, we will need to impose observational constraints post-hoc.

Real problems that deserve real resources

At the heart of the DestinE project are two real problems that do require increased focus from funders and researchers: the need for more direct (and more comprehensive) pipelines from climate projections to impacts, and the (related) need for a convergence of data of all kinds. This also relates to the push in the US to focus more funding on climate resilience and to better support climate policy decisions. A key step in these efforts is to be build platforms where all data sets (climate observations, reanalyses, projections, socio-economic data etc.) can be found (in the same accessible formats) and layered on top of each other at the appropriate scales. In DestinE, there is talk of a data lake, and efforts in the US to place more data in the cloud (e.g. Pangeo, EIS) go in the right direction. The challenges are finding formats that work in massively parallel data analytics platforms (zarr, xarray etc.), and in getting existing data centers to deliver in these formats. Given a full data platform, others can then build the interfaces that can answer the questions that any particular locality, sector, planner etc. may have.

DestinE data lake concept

Part of the answer will likely be developing open standards that will allow for a uniform query of multiple, structurally varying, (downscaled?) climate model output for each application so that updates, additions, and improved constraints, can be automatically passed down the pipeline. Additional effort will be needed to find ways to co-produce new models/assessments/projections that allow for feedbacks from the impact research to the upstream climate modeling efforts. This will not be easy.

Visualizations and Interactivity

The two user focused aspects of DestinE, are related to developing better visualizations and user interfaces (IMO this will be effort well-spent), and developing a new functionality for user input to projections such that they can easily see the impact of particular interventions. I have to confess that I do not quite understand what is envisaged here. Having users define changes in emissions and land use and having instant updates to projections will not be possible using actual climate models, so perhaps the idea is to use emulators informed by machine learning? If the interventions are on the impacts model side, it is perhaps easier to do, but the enormous heterogeneity of the impacts models will be the biggest challenge.

Digital Twin or Digital Evil Doppelgänger?

A digital twin for climate that allows us to skillfully try things out and have confidence that the real world will behave similarly would be a nice thing to have. A digital doppelgänger that only gives the impression of skill could be actively harmful. As always though, skill in any particular projections needs to be demonstrated, not merely assumed, so evaluations/hindcasts have to part of any new efforts. That means that the historical observations, reanalyses and climate model projections have to be seamlessly integrated so that reasonable assessments of forecast uncertainty can be made. Again, a very tall order.

Digital Twin-kle, little star, how I wonder what you are…

I don’t know what the DestinE project will produce: there are are clearly worthwhile aspirations built in to the plan, but they range from the tractable to the heroic. As other funders start to develop their own plans a close examination of which ideas are which will likely pay dividends.


  1. P. Bauer, B. Stevens, and W. Hazeleger, “A digital twin of Earth for the green transition”, Nature Climate Change, vol. 11, pp. 80-83, 2021.

  2. M.D. Zelinka, T.A. Myers, D.T. McCoy, S. Po‐Chedley, P.M. Caldwell, P. Ceppi, S.A. Klein, and K.E. Taylor, “Causes of Higher Climate Sensitivity in CMIP6 Models”, Geophysical Research Letters, vol. 47, 2020.

  3. C. Deser, R. Knutti, S. Solomon, and A.S. Phillips, “Communication of the role of natural variability in future North American climate”, Nature Climate Change, vol. 2, pp. 775-779, 2012.

Source link

Comment here

This site uses Akismet to reduce spam. Learn how your comment data is processed.