Plant ecosystems are highly sensitive to both short-term (seasonal) and long-term (climate change) atmospheric conditions. Historically, the representation of plant hydraulics in ecosystem modeling was limited, focusing primarily on soil moisture conditions. However, recent decades have seen significant advancements in plant hydraulics modeling to better simulate plant responses to varied soil and atmospheric conditions. This evolution has been driven by an increase in severe drought events, which have led to large-scale forest mortality, killing millions of trees over short periods. Plants employ diverse hydraulic strategies in response to drought and disturbances, reflecting their interspecies diversity. A trait-based approach has provided substantial flexibility in representing various plant hydraulic traits, though parameterizing these traits is complex and requires linking them to other measurable traits or environmental drivers. In response, we have developed a hybrid physics-informed machine learning (differentiable) plant hydraulic model, integrated within a broader differentiable ecosystem model. Using a machine learning component, we aim to learn and adapt to different plant hydraulic traits, thereby enhancing the simulation of plant transpiration by training the model using a large-scale sap-flow dataset covering multiple sites across the globe and extending over several years. This framework could potentially enable substantial improvement in our capability to learn plant hydraulic traits in trait-based approach models. It can also increase reliability for the study of hydraulic failures, moisture sensitivity, and vegetation responses to extreme droughts at large spatial scales. We discuss the impacts of solutions and differentiable programming techniques for large-scale learning.