Towards a Hybrid Vegetation Model

Vegetation plays a crucial role in the Earth's climate system via a number of key processes, affecting the exchange of carbon, water and energy between surface and atmosphere. The complex relationship between climate change and vegetation highlights the importance of accurate and reliable vegetation models that fully capture these interactions. 

Traditional vegetation models are primarily designed to operate on CPU architectures, which restricts their ability to exploit advancements in modern parallel computing architectures such as GPUs. Furthermore, limited process knowledge and the absence of direct observations and/or quantitative theories for certain processes hinder accurate representation of these processes, which introduces uncertainties in the model results, leading to discrepancies when compared to observations. The rigid structure of these traditional models also makes integration of new processes challenging and hinders the application of advanced optimization techniques for automatic parameter tuning and objective calibration using abundant observational data due to their non-differentiable nature.

This work follows a new paradigm in vegetation modeling that integrates the robustness of traditional models with the adaptive power of machine learning techniques. The goal is to combine reliable physical components with machine learning components. As opposed to classical vegetation models, the resulting hybrid model is differentiable and the parameters of both the physical and the neural network components can be optimized jointly and efficiently using observational data.

In the proposed hybrid vegetation model, machine learning can be used to improve the computational efficiency of the model by emulating computationally expensive routines. We have implemented the key processes related to photosynthesis in LPJ in Julia. This minimal model setup is used to explore the potential of machine learning to replace the computationally expensive root-finding algorithm used in computing the optimal ratio of intercellular to ambient CO2 concentration, and hence stomatal conductance.

Machine learning can also be used for better process representation. The recently developed neural or universal differential equations offer a particularly promising methodological framework for learning the dynamics of carbon allocation to different vegetation pools using observations. The dynamics of carbon allocation to different plant components can be effectively modeled using a neural ODE approach, which utilizes observations of observable variables (e.g., Above Ground Biomass (AGB), Leaf Area Index (LAI)) to learn the dynamics of unobservable variables such as vegetation carbon pools.