Net photosynthesis (A_N) is the largest flux of the global carbon cycle, shaping the decadal-scale climate change. Present earth system models (ESMs) models estimate A_N using various parameterization approaches for different photosynthesis parameters, on top of which the maximum carboxylation rate (V_c,max25). Some ESMs adopt plant specific dependent parameters while others expand the acclimation of these parameters to include the plant physiological characteristics or the surrounding environmental conditions. While vegetation acclimation to the  surrounding environment has been deemed important, it was challenging to incorporate such knowledge with good spatial generalizability. Here we used a differentiable (physics-informed machine learning) model to test various parameters’ acclimation approaches. We used neural networks to learn acclimation functions for different parameters of the Farquhar-type photosynthesis modules, and trained them on multi source dataset including A_N, stomatal conductance (g_s) and V_cmax25 data simultaneously. Our model generalized well in space better than traditional and alternative models showing a great potential to learn from multi-source datasets. The model simulations show that the acclimation approach can impact our estimations for the global net photosynthesis. These results have strong implications for climate projections.