Higher-level geophysical modelling

Until recently, our pure Python, primitive equation ocean model Veros 
has been about 1.5x slower than a corresponding Fortran implementation. 
But thanks to a thriving scientific and machine learning library 
ecosystem, tremendous speed-ups on GPU, and to a lesser degree CPU, are 
within reach. Leveraging Google's JAX library, we find that our Python 
model code can reach a 2-5 times higher energy efficiency on GPU 
compared to a traditional Fortran model.

Therefore, we propose a new generation of geophysical models: One that 
combines high-level abstractions and user friendliness on one hand, and 
that leverages modern developments in high-performance computing and 
machine learning research on the other hand.

We discuss what there is to gain from building models in high-level 
programming languages, what we have achieved in Veros, and where we see 
the modelling community heading in the future.