Towards ecologically meaningful foundation models

Ecology seeks to explain and predict interactions among organisms and their environments across space and time, yet both data and theory remain fragmented. Empirical evidence spans diverse modalities, scales and contexts, while theoretical frameworks are rarely integrated into unified predictive models. This paper proposes ecological foundation models (ecoFMs), trained on large multimodal datasets, as a path to unify data and theory. EcoFMs could learn shared representations of organisms, environments and interactions, improving generalisation, linking pattern to process and enabling synthesis across sub-disciplines. They also provide a demanding testbed for machine learning, requiring advances in multimodal representation learning, theory-guided modelling and inference.