The rapid growth of remotely sensed earth observation data presents clear opportunities for monitoring complex ecosystem change and answering fundamental ecological questions. However, large-scale automated monitoring of ecosystems faces challenges. Data-driven models require extensive datasets and often lack generalizability when training data are unrepresentative, while process-driven models, such as radiative transfer models (RTMs), can be imprecise due to gaps in knowledge, or simplified representation of physical processes. To enhance the prediction of plant functional traits and simultaneously discover where process-driven models can be improved, we explore the potential of Physics-Informed Neural Networks (PINNs) as a hybrid approach that combines the strengths of both methodologies at the leaf scale. In contrast to data augmentation approaches, our implementation directly integrates the widely-used PROSPECT5B model into the architecture of an autoencoder framework. Our results show that our PINNs approach is able to outperform data-driven techniques even when trained on very limited training data (i.e. 17 % training vs 83 % validation). We also identified weak points in the PROSPECT5B model by progressively replacing individual components of PROSPECT5B with convolutional neural networks. Our case study indicates that especially Prospect's generalized “plate model” could be refined to improve predictive ability. Hence, our framework provides a self-diagnostic capability and identifies areas for improvement in process-driven models and their components. Thus, we conclude that PINNs 1) improve data-driven predictive accuracy while maintaining physical consistency with minimal training data while 2) being able to identify limitations in process-driven models. Hence, we believe our framework could serve as a new standard for evolving and improving radiative transfer models.