A novel computational framework for optimal experimental design to improve climate prediction

There is a pressing but unmet need to optimize the deployment of climate observing systems. Many observing systems are expensive and require extensive planning before finalizing the design. Resources are often misallocated, and critical insights are not identified during the lifetime of the observing systems. Therefore, we are developing a novel computational framework to improve model predictability by integrating key components: model simulation, uncertainty quantification, and optimal experimental design. Our study specifically utilized the DOE Energy Exascale Earth System Model (E3SM) land model (ELM) to simulate land-atmosphere carbon, water, and energy flux quantities that are critical to study biosphere-atmosphere exchange. We investigated 26 parameters that regulated vegetation structure and function based on domain knowledge and used the Sobol sequence to generate a quasi-random space-filling parameter set. An ensemble simulation with 1300 ensemble members was performed for each site. Then we used a statistical emulator as a surrogate of ELM to reduce the computational costs of carrying out model-data fusion efforts and model calibration. Bayesian probabilistic parameter estimation was applied to the emulator wherein posterior probability distributions on model parameters are updated based on the available field observations. Finally, an observing system simulation experiment (OSSE) was developed and tested with existing data to guide future deployment strategies.