Physics-informed Recurrent Neural Network Surrogates for E3SM Land Model

Complex climate models such as Energy Exascale Earth System Model (E3SM) are computationally expensive. This leads to challenges associated with ensemble-intensive studies such as parameter estimation, uncertainty quantification and optimal experimental design. In order to make such studies more tractable, we build efficient and accurate surrogate approximations to functional maps from input parameters to output quantities of interest (QoIs).

This study focuses on building neural network surrogates for the ELM, the land model component of the E3SM. Due to the temporal nature of the model, we construct the surrogate with recurrent neural network (RNN) architecture with gated units, e.g. long-short term memory (LSTM), to enable long-range propagation of information for daily QoIs such as total leaf area index (LAI) and gross primary productivity (GPP).

Furthermore, we augment the architecture with a hierarchical structure of the relationships between the various quantity of interests (QoIs) as well as the physical processes. The resulting hierarchical LSTM RNN mimics the physical constraints and relationships between the underlying processes of phenology, respiration and allocation, and provides a natural `prior' for temporal evolution. The physics-informed hierarchical LSTM surrogate outperforms conventional LSTM, multilayer perceptron and polynomial expansion surrogates in terms of accuracy. Using the hierarchical LSTM surrogate, we then perform global sensitivity analysis to identify the most influential input parameters for dimensionality reduction. The surrogate is planned to be employed further for model calibration given sets of observational data at FLUXNET sites.