# Feedback Lily's concept defense

All three advisors agree that you are well prepared for the rest of the PhD. You have realized very important things over the past months (even if you don't always realize it):
- the suite of MB models in OGGM-sandbox is a fantastic addition, and the snow-aging model is unprecedented in the large-scale community
- the re-implementation of the Bayesian calibration (including new additions such as PyMC3 or the new prior estimation) is a difficult task that you mastered well
- you have made several contributions to OGGM core already
- your participation in GlacierMiP3 is an excellent development for your career and networ
- your co-authorship on Sarah's paper in progress is also very fruitful
- and more!

We have no doubts that you are on a good track for your PhD!

The main challenge ahead is that you probably already have more results than one can chew. You are a very meticulous and perfecionist scientist, which is a blessing but also a curse. All that new information each day / week (also coming from me as an advisor!) can become overwhelming. It may happen that you ten to see only problems ("are we really allowed to do that?"), blurring the way to publication.

Our task for the next months (until the submission) will be to clear the bushes and focus on parts of the problem and discard or ignore the others in order to simplify the message of the paper.

Based on our discussions, here are some suggestions to consider for paper 1:
- "Part A" could serve the purpose of (i) demonstrating that the equifinality problem is the same for all kind of models and (ii) pick the one model to take to part B.
- for (i), we show all the differences you showed already between the models. We will do the tests with fixed geometry to disantangle dynamical effects (as in [this paper](https://www.frontiersin.org/articles/10.3389/feart.2017.00056/full) I think)
- for (ii), we should take the one which provides "the best" results on the reference glaciers, with respect to traditional metrics of SMB timeseries, and wrt the MB gradients (profiles? ;-). The other models are discarded, but we will argue that the methods developed in part B are applicable to all models.
- A great achievement of Part A will be that these models are now available for all users in the OGGM framework.
- "Part B" will try to address one question at first: can we disantangle uncertainty from equifinality from uncertainty of observations? We use the Bayesian framework for this, some ideas have been mentioned at the meeting. If all goes well, we should be able to do this at the glacier scale.
- For aggregating results at the regional scale, we need methods that we don't have yet. Nobody has them. We all agree that it would be a major step forward to implement even parts of a solution for uncertainty propagation into OGGM. Our suggestion is to continue to think about this problem of the next few weeks (~ 4) and see if you can make progress. I will help to draft an email to some prominent glacio-mathematicians (Alex Robel, Andy Aschwanden, Lizz Ultee, Doug Brinkerhof, any other? *Lindsey works with some pretty interesting UK mathematicians (Geoff Evatt/Matthias Heil/Dave Abrahams ... and could maybe interest them in this, though not sure of the nature of the problem really fits in their wheelhouse, but they like working on glaciology - especially Geoff! Could ask at least!).*

Even with only a partial implementation of the uncertainty propagation part, the paper would be a great method paper suitable for JAMES or GMD.

Regarding our work together over the next months, I suggest to include Dave on a more regular basis (every two weeks?) and Lindsey when needed (month? 2 months)? The slack channel for Lily will continue to be #bayesian.

I know that I (Fabien) can sometimes have a lot of (contradicting) ideas, and the regular check ups with both of the other advisors can help to re-focus a little.

I (Lindsey) would also bring up the question of whether the sandbox itself is worthy of a GMD paper, which might also help streamline this first real science paper as you can refer to the modelling tools published elsewhere? [Here are my own HackMD notes](https://hackmd.io/z0GSiOH8RD2C8xqp3EtAfw?view) on your text which will let you see a bit about what was clear/not clear to me! I hope as a somewhat outside eye on your work I can support you in making choices and priorities for carving out achieveable and satisfying publications.

David adds:
- In the thesis, there was a lot of discussion on only have two parameters (precipitation factor and melt factor). This is fine and good for “good” glaciers, which perhaps the reference glaciers all are; however, when you go to regional scales, you’re going to have to include a temperature bias. Therefore, I’d be cautious with how the model description and also with going too far with the use of precipitation data to limit the precipitation factor (an idea that was in the .pdf but perhaps excluded) because ultimately the precipitation factor (and melt factor) will be compensating for everything else that’s missing.
- *Fabien: OK - we actually discussed this at length with Lily. Let's try without first, and then see if we can generalize to 3 parameters.*
- For non-reference glaciers that lack data, as more snowline altitude data becomes available (or perhaps it already is in specific regions), you could potentially use the interannual variations in the snowline altitude as a proxy for the interannual mass balance.
- *Fabien: yes totally, but I'm a bit scared of opening this rather large WP. If the uncertainty propagation doesn't work at all, we might get back to this idea.*

**Would that sound like a plan, Lily?** Is there anything you would like to say to us as well regarding feedback, supervision, etc?