Machine Learning for Earth Systems Modeling, Analysis and Predictability

Artificial intelligence (AI) and machine learning (ML) methods and applications have been continuously explored in many areas of scientific research. While these methods have lead to many advances in climate science, there remains room for growth especially in Earth System Modeling, analysis and predictability. Due to their high computational expense and large volumes of complex data they produce, earth system models (ESMs) provide an abundance of potential for enhancing both our understanding of the climate system as well as improving performance of ESMs themselves using ML techniques. Here I demonstrate 3 specific areas of development using ML: statistical downscaling, predictability using non-linear latent spaces and emulation of complex parametrization. These three areas of research illustrate the ability of innovative ML methods to advance our understanding of climate systems through ESMs.

In Aim 1, I present a first application of a fast super resolution convolutional neural network (FSRCNN) based approach for downscaling earth system model (ESM) simulations. We adapt the FSRCNN to improve reconstruction on ESM data, we term the FSRCNN-ESM. We find that FSRCNN-ESM outperforms FSRCNN and other super-resolution methods in reconstructing high resolution images producing finer spatial scale features with better accuracy for surface temperature, surface radiative fluxes and precipitation.

In Aim 2, I construct a novel Multi-Input Multi-Output Autoencoder-decoder (MIMO-AE) in an application of multi-task learning to capture the non-linear relationship of Southern California precipitation (SC-PRECIP) and tropical Pacific Ocean sea surface temperature (TP-SST) on monthly time-scales. I find that the MIMO-AE index provides enhanced predictability of SC-PRECIP for a lead-time of up-to four months as compared to Ni{\~n}o 3.4 index and the El Ni{\~n}o Southern Oscillation Longitudinal Index. I also use a MTL method to expand on a convolutional long short term memory (conv-LSTM) to predict Nino 3.4 index by including multiple input variables known to be associated with ENSO, namely sea level pressure (SLP), outgoing longwave radiation (ORL) and surface level zonal winds (U).

In Aim 3, I demonstrate the capability of DNNs for learning computationally expensive parameterizations in ESMs. This study develops a DNN to replace the full radiation model in the E3SM.