That’s wild that y’all only implemented this recently

Agree that emulators are great too, especially since we can do that now. I would love to see deep learning approaches used to increase ensemble sizes of climate models. Would be incredible to have 3000 member ensembles of a few select variables like temperature and precip

Optimal critical humidity is just an example parameter I pulled from your paper describing ModelE. In theory, you could do this with every single parameter if you wanted to

This is a massive software engineer task for something like ModelE. Full rewrite of the model to make the model differentiable (using a library like JAX). Additionally (as with manual tuning) maintaining numerical stability, and stable climate states, can be very challenging

Different approach here, the goal is to learn the parameters of a model directly from observations rather than manual tuning. Model could be simple or complex like ModelE, but with a differentiable model, the optimal critical relative humidity would be found by gradient descent

There's a way forwards though, which is very exciting. For example, you could include a radiative transfer model of your choosing in this framework

Something like NeuralGCM, which combines a dynamical core with a learned subgrid physics module generalizes pretty well

www.nature.com/articles/s41...