UC Irvine

Ecosystem dynamics are heavily dependent on atmospheric inputs such as rainfall, and are in turn an integral part of land-atmosphere coupling and the global carbon cycle. These global interactions and cycles are commonly modeled by Earth System Models (ESMs), and one of the largest sources of uncertainty in current models is cloud simulation techniques.

Superparameterization (SP) is a proven method to better resolve cloud and rainfall processes in ESMs, but it is prohibitively computationally expensive for long-term climate simulations. The first part of this thesis suggests that – although the latest advances in manycore supercomputing do not provide a promising solution to this problem – a neural network (NN) can successfully be trained on an SP version of the Community Atmosphere Model (CAM) to emulate SP behavior at a fraction of the cost. Incorporating the NN into CAM under idealized aquaplanet conditions results in the Neural Network Community Atmosphere Model (NNCAM).

For NNCAM to work in fully comprehensive models that include interactive vegetation, deeper tests are needed. The second and most major contribution of this thesis is thus to analyze the one-way coupling of NNCAM onto the Community Land Model (CLM). Gross primary productivity (GPP) and net ecosystem exchange (NEE) from ensembles of one-way atmospheric forcing onto a CLM grid cell in the Amazon Basin for NNCAM and SP-CAM are shown to be statistically different from CAM but not each other. Additionally, results suggest that mean precipitation is the largest contributing factor to GPP and NEE in the Amazon Basin.