Inferring processes within dynamic forest models using hybrid modelling

Abstract Modelling forest dynamics under novel climatic conditions requires a careful balance between process‐based understanding and empirical flexibility. Dynamic vegetation models (DVM) represent ecological processes mechanistically, but their performance is sensitive to misspecified functional forms and to unavoidable structural simplifications. Inferring the structure of these processes and how they best align with each other within a DVM from data remains a major challenge because current approaches, such as empirical plug‐in estimators, have proven ineffective. Here, we introduce Forest Informed Neural Networks (FINN), a hybrid modelling approach combining a forest gap model and deep neural networks (DNNs). FINN embeds DNNs instead of mechanistic process formulations, enabling the DNNs to learn the most effective predictive process forms while aligning them with the other processes. FINN accomplishes this by calibrating the embedded DNN jointly with the other components of the DVM. FINN is implemented in torch for R, a deep learning framework that makes it fully differentiable and enables efficient, end‐to‐end optimization. Using simulated data, we demonstrate that this end‐to‐end optimization reliably recovers process parameters and known functional forms. In a case study of the 50‐ha Barro Colorado Island plot, we replaced the growth process with a DNN and found that, while the process‐based and hybrid models had similar short‐term predictive performance, only the hybrid model produced ecologically plausible long‐term succession trajectories; a naïve DNN without process scaffold failed to produce plausible dynamics despite accurate short‐term rate predictions. Using explainable AI, we found that the DNN learned an ecologically plausible functional form of growth that differs from its mechanistic counterpart in interpretable ways. In conclusion, FINN demonstrates that a hybrid approach can improve the internal consistency of DVMs not only by learning the best predictive form from data but also by aligning processes in a way that leads to reliable long‐term trajectories. This paves the way for hybrid modelling, which allows us to infer forest dynamics from data and improve forecasts of ecosystem trajectories under unprecedented environmental change.