Laying the foundations for a hindcast

Predicting the past

This week I wrote the words “finish thesis draft” in my one-year plan, which means that it is T minus 12 months until the end of my PhD. It’s scary stuff, but also exciting, because I am now entering the final, and most exciting, stage of my research. The last phase of my work involves using what I’ve already done on wintertime foehn-driven melt and cloud microphysics to optimise a model configuration and then use this to produce a long-term hindcast. A hindcast is like a forecast – using a model to project future conditions – except that we do the same thing for the past. Typically, a hindcast is compared with available data to make sure it’s doing something sensible, and then it can be used to see what conditions were like during times and at locations where we don’t have observations.

In Antarctica, a model hindcast is a particularly useful tool. Because it is so difficult to make observations in the harsh Antarctic environment, we are limited to conducting measurement campaigns during the summer. While there are a few automatic weather stations that record throughout the year, most of these have only been deployed within the last decade, so it’s hard to see trends and patterns yet in the data. This means that we don’t have a great handle on what is going on during winter, nor on how things have changed.

So: a model hindcast is the next-best thing. If we can be confident that the model does a reasonable job at representing conditions during the times when we have observations available to compare it to (i.e. that we can validate it against), then we can assume that it will do a similarly good job during times when we don’t have data.

Following the work I did finding the best model configuration for representing cloud properties, I have arrived at a least-worst combination of physics settings that get average cloud fields pretty well over Larsen C. I’m happy that I can use this set-up and it will do a good enough job.

What’s exciting about this project is that it hasn’t been done before – while a climatology of wind and temperature over Larsen C has been done with a model called RACMO, a comprehensive hindcast of cloud, surface energy balance (SEB), surface mass balance (SMB) and meteorology hasn’t. What is more, this previous work used a hydrostatic model, which is incapable of representing foehn dynamics because of the assumptions made in the core physics (the hydrostatic assumption) and its necessarily coarser horizontal resolution (5.5 km). Foehn dynamics have been shown by countless papers and my own work to be very important in controlling the weather, climate and melt regimes over Larsen.

This hindcast will be a great resource: providing a time series of many of the factors controlling melt over a longer time period than we have observations for. I will be using the ERA-Interim re-analysis dataset (our best guess of what the atmosphere was doing at any point since 1979, produced from all available observations) to force the model, initially for the most recent decade, but with a view to extending this to the start of the satellite period in 1979. From this, we can establish a baseline of what conditions are like, and therefore start to understand the changes that we see – and potentially use this to project into the future.


Realistically it’s quite time-consuming to run a multi-year simulation in one go, so I’m splitting it up into one-year chunks. What I will do is use the model in forecast mode, which is the operational configuration that the Met Office uses to forecast the weather. Every twelve hours the model is fed re-analysis data from ERA-Interim. Once it’s set off with this data, it is allowed to run forwards and calculate conditions for the next 24 hours. I’m discarding the first 12 hours of each forecast (because the quality of the forecast is best between 12 and 24 hours) and then stitching together the 12-24 hour period of each to make one continuous time series.

In theory, many of these one-year chunks can be run at once, and then combined into one long hindcast. It also means that I can keep adding years, stretching backwards in time, to produce as long a hindcast as I can (although the time limitations of my PhD will probably be the main reason that I have to stop adding years!). I want to have at least ten years to analyse, spanning the period 2008-2017, during which we have the most abundant observational data.

But why?

That’s all well and good, but there’s no point in creating a huge model dataset without having an idea of the kind of questions that you want to answer using it. The model set-up has therefore been designed with some specific themes and ideas in mind, which build on the work I have previously done looking at the effect of foehn winds and cloud microphysics on the surface energy balance of the ice shelf. I want to understand the effect of atmospheric processes on the amount of energy received at the surface (the SEB), and extend this to find out what effect this has on Larsen C’s surface mass balance. The mass balance is another one of those accounting terms that describes the relationship between how much ice goes in vs how much comes out. From the SMB we can figure out whether Larsen C is growing or shrinking, and from that begin to predict its future.

Specifically, I’m interested in finding out how frequently foehn events occur, and what effect they typically have on melt rates. We know that foehn influence surface melting, and are indeed responsible for producing melt ponds at the surface in inlets like Cabinet Inlet (satellite pic), but can we use the hindcast to understand exactly how foehn events drive this melt, and whether this varies between seasons? What is the fate of the meltwater produced? Does it refreeze, is it contributing to firn densification and ultimately hydrofracturing processes, or does it run-off?

And what about clouds? Although the model’s representation of cloud microphysics leaves something to be desired, can we say something about the role of clouds in driving melt? For instance, how does cloud composition (especially the amount of liquid and ice) affect the amount of energy that reaches the surface, and therefore the amount of meltwater that is produced? Are clouds a dominant factor in determining how much the ice shelf beneath it melts? And if so, does that give them significant control over the SMB?

Broadly: when do we see the most significant melting, and what kind of conditions are dominant when this is observed? Are certain types of weather systems more likely to set up melting? And, how important is surface melting anyway in determining the overall SMB of Larsen C?

These are all interesting questions, and no doubt I will think of more as I go along. I’d love for others to be able to use the hindcast once it’s put together to answer their own burning science questions, so get in touch if that’s you. But for now, I will return to gazing backwards through my crystal ball…


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s