Net photosynthesis (AN) is a major component of the global carbon cycle, with significant feedback to decadal-scale climate change. Although plant acclimation to environmental changes can modify AN, traditional vegetation models in Earth System Models (ESMs) often rely on plant functional type (PFT)-specific parameter calibrations or simplified acclimation assumptions, both of which lacked generalizability across time, space and PFTs. In this study, we propose a differentiable photosynthesis model to learn the environmental dependencies of Vc,max25, as this genre of hybrid physics-informed machine learning can seamlessly train neural networks and process-based equations together. Compared to PFT-specific parameterization of Vc,max25, learning the environment dependencies of key photosynthetic parameters improves model spatiotemporal generalizability. Applying environmental acclimation to Vc,max25 led to substantial variation in global mean AN, calling for the attention to acclimation in ESMs. The model effectively captured multivariate observations (Vcmax25, stomatal conductance gs, and AN) simultaneously and, in fact, multivariate constraints further improved model generalization across space and PFTs. It also learned sensible acclimation relationships of Vc,max25 to different environmental conditions. The model explained more than 54%, 57% and 62% of the variance of AN, gs, and Vcmax25, respectively, presenting a first global-scale spatial test benchmark of AN and gs. These results highlight the potential of differentiable modeling to enhanced process-based modules in ESMs and effectively leverage information from large, multivariate datasets.