Ed Gerber

Ed Gerber

Professor at the Courant Institute

Emulating state-of-the-art gravity wave parameterizations

less than 1 minute read

Published:

Don’t miss Y. Qiang Sun’s paper on the use of machine learning to emulate a physics-based atmospheric gravity wave parameterizations in a state-of-the-art climate model, in review in JAMES. It’s also up on the arxiv.

This is a tour-de-force study, where Qiang emulated orographic, frontal, and convective gravity wave parameterizations in WACCM, which has perhaps the most sophisticated treatment of gravity waves on the market! He first had to overcome set imbalance to make training possible, and then probed uncertainty in the emulators, showing that many of the current approaches provided equivalent information. Finally, he showed that while the machine learning based emulators couldn’t generalize out-of-set to climate perturbations, with just a bit of data from the new climate, transfer learning can fix a scheme. This suggests that a small amount of data from high resolution simulations in warm climates might be sufficient for the purposes of data-driven parameterizations trained on today’s climate.

You May Also Enjoy

See the forest and the trees!

1 minute read

Published:

Congratulations to Dave Connelly, who just submitted his first paper! It is about the use of regression forest to represent atmospheric gravity wave momentum transport to JAMES! The manuscript makes two important steps forward. First, it shows that a “boosted forest” approach, where you train each subsequent decision tree on the residual (as sketched below), can out perform a “random forest” where you combine a number of decision trees, averaging the result. This was well known in the ML community, but less so in the climate sciences. Second, Dave found that techniques from interpretable AI could be used to improve the training of a data driven parameterization. Using feature importance metrics, he found that his origional boosted forest wasn’t using enough information about latitude. By forcing the method to predict the latitude as well, he could build trees that incorporate this information more effectively!

The graft-versus-host problem…

less than 1 minute read

Published:

Ofer Shamir just submitted a paper on the use of machine learning to represent atmospheric gravity wave momentum transport to QJRMS. He developed an idealized, one-dimensional model of the Quasi-Biennial Oscillation, where we could systematically compare different data-driven methods. In particular, how can one calibrate a data-driven scheme to work in a biased model (the graft-host problem), and how well can schemes generalize to new conditions, as in a climate change scenario?

Um, what happened to me?

less than 1 minute read

Published:

I flatter myself to think you may have stopped by my webpage over the last year and half and wondered, what happened to Ed? No new papers? No research? Did he drop off the face of the Earth? Sort of. I was on sabbatical at Free University Berlin and Ludwig Maximillian University, Munich for an academic year!