Adam Michael Erickson , Craig Nistchke 

Soil moisture is a key limiting factor of plant productivity in boreal and montane regions, producing additional climate feedbacks through evaporation, regeneration, mortality, and respiration. Understory solar irradiation – the primary driver of surface temperature and evaporative demand – remains poorly represented in vegetation models due to a lack of 3-D canopy geometry. Existing models are further unable to represent processes lacking sufficient parameterization and/or knowledge, with no land model to date utilizing machine learning (ML) to represent vegetation processes. Here, we developed the first hybrid forest ecosystem model using ML (ML-FEM), a specific case of hybrid AI land model (a concept also invented here). In this approach, ML models are trained and validated with a ground-truth dataset, whether observations or high-fidelity simulations, before being applied to vegetation model parameters for inference, internally or externally to the model. Using this approach, we simulated annual understory global solar irradiation (Iu) across 25.2 Mha in southwestern Canada at 1-ha resolution under historical climate and fire scenarios. In cross-validation, we found that linear and ML regression models performed comparably well in the prediction of angular canopy cover (ACC), due to the linearity of its relationship to predictors (linear R2 = 0.938, RMSE = 0.079; ML R2 = 0.939, RMSE = 0.074). Reduced area burned, increased ignitions, and reduced regeneration potential for recent periods resulted in stable or reduced Iu. This suggests that diminished disturbance may reduce Iu through forest aging, masking latent regeneration decline. Only in the most extreme and unconstrained scenarios did Iu increase. In these experiments, conducted in late 2015, we demonstrated an entirely new class of hybrid models that we anticipated to be of vital importance to understanding and representing pattern-based processes in Earth system models.