Home

Rasp, Gentine 2018: Deep learning to represent subgrid processes in climate models

The computational expensive cloud resolving model, SPCAM, is surrogated by training a deep neural network (NNCAM).

• Once trained, the NNCAM runs 20 times faster

• The NNCAM captures the mean climate well

• NNCAM learned to conserve energy

• The network is able to interpolate between different temperature scenarios but fails on extrapolation

• NNCAM outperforms a parametrized surrogate model, CTRLCAM

Method:

• NN with 9 layers and 256 nodes per layer

• dimension of data point: x=94, y=65 with 140 mio training points

Question and Discussion:

• Probabilistic surrogate model, e.g. GP, include prior knowledge

• from qualitative to quantitative uncertainty or error estimation

• real world data instead of simulations, why are sparce or missing data a problem?