Process-based Neural Network to Forecast Vegetation Dynamics

Abstract

Process-based Neural Network to Forecast Vegetation Dynamics Chonggang Xu, Youzuo Lin, Nishant Panda, Monty Vesselinov, Humberto Godinez Vazquez Los Alamos National Laboratory Focal Area 2 This white paper calls for an AI-based emulator of DOE’s dynamic vegetation model that is computationally efficient and accommodates/parameterizes the governing processes built in the model. This AI-based emulator is needed for model calibration, sensitivity analysis, prediction, uncertainty quantification, and decision making. This AI-based emulator will incorporate physics and biological information in the form of conservation laws and constitutive relationships. Science Challenge DOE’s demographic vegetation model, the Functionally Assembled Terrestrial Simulator (FATES), allows comparison with many more observable vegetation processes than the first generation ‘big leaf’ vegetation models, but also faces two key challenges. First, FATES contains more degrees of freedom leading to greater complexity and more uncertainties in the parameter estimations. Second, because more processes (e.g., vegetation recruitment, growth and mortality) are introduced, FATES is computationally much more expensive. These two challenges make it difficult to calibrate the model against observations using traditional approaches for parameter estimation (e.g., maximum likelihood or Markov Chain Monte-Carlo). One solution is to build emulators to efficiently mimic these demographic vegetation models; however, due to the larger number of parameters and model complexity, it is only feasible to build emulators with a constrained and limited number of key parameters based on traditional kernel methods 1. Rationale Earth System models (ESMs) are evolved from the general circulation models (GCMs), which mainly focus on the physical process of ocean and atmospheric circulations, with a new focus on the climate feedbacks from biological systems2. An important biological component is vegetation, which plays a critical role in global and regional water cycles3. Specifically, the reduction in plant productivity and increasing vegetation mortality resulting from extreme climate conditions could substantially affect the regional and global water cycles 3,4. The first-generation vegetation models represent plant communities by area-averaged leaf layers of different plant functional types (PFTs) within each land grid cell. However, these bigleaf vegetation models do not represent the coexistence of different sizes of plants or PFTs, meaning that these models cannot express the competition for light, water, and nutrients among plants. To overcome these limitations, scientists have incorporated ecosystem demographic models into ESMs5,6. These new models include an ecosystem’s demographics by explicitly simulating plant size, diameter growth, mortality, and recruitment based on competition for light, nutrients, and water5. Because these demographic vegetation models explicitly represent the coexistence and competition among different size groups of PFTs, they are expected to better represent changes in regional and global water cycles associated with disturbances. Specifically, DOE’s FATES is considered a next-generation vegetation model for E3SM, with a size-structured group of plants (cohorts) and successional trajectory-based patches using the ecosystem demography approach7. Within FATES, the transpiration rate is mainly determined by leaf-level photosynthesis, competition as determined by growth and mortality, and hydraulic function changes under water stress. FATES allocates photosynthetic carbon to storage and different tissues, such as leaf, root, and stem based on the allometry of different plant species. Mortality within FATES is simulated by carbon starvation caused by depletion of carbon storage