Photosynthetic simulations in Earth system models (ESMs) typically rely on some stationary or empirically derived parameters. This is because previous vegetation modules in ESMs either calibrated parameters specific to each plant functional type (PFT) or used simplified assumptions for plant acclimation to environmental conditions. Neither approach adequately addresses the impacts of acclimation, resulting in poor spatial generalizability and trade-offs between simulated variables. To address these limitations, we employed a hybrid physics-informed machine learning (differentiable) ecosystem model. This model learns the environmental dependencies of photosynthetic parameters from multi-source datasets, greatly improving model generalizability and potentially revising global photosynthesis estimates. Additionally, the model can capture multiple objectives simultaneously. Our differentiable models identified sensible relationships between acclimation-related parameters, such as Vc,max25, and various environmental conditions. The results indicated that global net photosynthesis could be overestimated when heterogeneous data sources were not utilized, or when environmental acclimation of parameters was not considered, as is common in traditional approaches. For a better representation of plant hydraulics, we are working to leverage the flexibility of our differentiable ecosystem model and replace the empirical physical equations based solely on soil moisture conditions with a differentiable plant hydraulic component based on the hydraulic module incorporated into the Functionally Assembled Terrestrial Ecosystem Simulator (FATES-HYDRO). By integrating a plant hydraulic component, we aim to better learn different plant hydraulic traits by linking them to measurable plant traits using a machine learning model. Such an approach makes the framework applicable for studying the plants’ hydraulic failures and their sensitivity to various drought conditions.