Work

Inferring Plant Acclimation and Improving Model Generalizability With Differentiable Physics‐Informed Machine Learning of Photosynthesis

Year: 2025

Type: article

Abstract: Differentiable ecosystem model (δpsn) ramework. The δpsn framework integrates a neural network (NN) component with a differentiable process‐based model in this case the photosynthesis module from FATES. The NN component includes three networks—NNB, NNV, and NN⍺—each trained on different sets of predictors, such as soil attributes, plant functional type (PFT), leaf nitrogen content (LNC), and environmental conditions. These networks learn distinct parameters (including Vc,max25—the maximum carboxylation rate at 25°C, a key parameter representing plant photosynthetic capacity—and soil water constraint parameter Bi) that feed into the photosynthesis module along with meteorological forcings (F) and constant terrain attributes (θc). NN⍺ captures parameter acclimation to environmental factors, and predicts a perturbation term (⍺p) which is multiplied by each parameter (p) learned by NNV. The photosynthesis module then simulates AN (net photosynthesis) and gS (stomatal conductance). The outputs, including Vc,max25, are compared against multivariate observations to compute a loss function, which guides the backpropagation process to update the NN parameters.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.15691215