Thursday, March 19th, 2026
Marine socio-ecological systems are profoundly interconnected. Species interact through food webs, habitats influence community structure, and human activities shape ecological outcomes in ways that ripple across space and time. Traditional modelling approaches capture some of these relationships, but the growing scale and complexity of marine data create new challenges. Graph Neural Networks offer a promising response. The EcoTwin report, Preliminary report on framework of causal models, presents a clear, forward-looking explanation of how these deep learning models can contribute to causal discovery in marine environments.
The report introduces Graph Neural Networks, or GNNs, as a family of machine learning models designed to operate on graph-structured data. Graphs consist of nodes and edges, and represent relationships between entities such as species, habitats or socioeconomic factors. The report begins this explanation by outlining graph fundamentals, including adjacency matrices, node features and the graph Laplacian. These mathematical tools allow GNNs to analyse patterns in complex networks without reducing them to simple, independent variables.
The authors explain that GNNs work by passing information along edges so that each node updates its internal representation based on its neighbours. This is known as message passing. In ecological terms, this is a natural fit. Species are influenced by the species they interact with. Habitats are influenced by the habitats they are connected to. Human communities respond to the behaviour of neighbouring communities or fleets. GNNs are designed to learn these patterns directly from data, allowing them to capture relationships that are difficult to encode using traditional modelling techniques.
After introducing the basic architecture, the report surveys several important GNN variants. These include spectral convolutional networks, which use the eigenvalues and eigenvectors of the graph Laplacian, and Graph Convolutional Networks, which apply localised operations that generalise classical convolution to networks. The report also discusses Message Passing Neural Networks, which unify many GNN formulations under a single framework. This overview is presented in the same section.
In a discussion of how GNNs relate to causal inference, the EcoTwin authors describe a conceptual link between the structure of a causal graph and the architecture of a GNN. They suggest that it is possible to construct a GNN whose graph matches the underlying causal structure of a system. In this approach, the causal graph determines how messages flow between nodes, allowing the GNN to approximate the behaviour of a Structural Causal Model.
This idea opens several possibilities. If the causal graph is known or partially known, a GNN can be designed to respect those dependencies, ensuring that learned relationships are consistent with known mechanisms. If the causal graph is unknown, the GNN may help identify which relationships are most likely to represent causal pathways. In this way, GNNs can support both causal modelling and causal discovery.
The report further explains that GNNs can be extended to handle interventions. When a variable is manipulated in a causal model, its relationships with parent variables change. The report describes how similar changes can be applied within a GNN. By modifying node states or message passing rules, it is possible to simulate interventions and observe how changes propagate through the network. This creates an opportunity to use GNNs for estimating causal effects.
Beyond these structural links, the report introduces several advanced techniques for integrating causality with deep learning. These include interventional Graph Convolutional Networks, which explicitly represent the effects of interventions, and variational graph autoencoders, which can learn latent structures that approximate causal relationships. The authors also reference work on universal density approximation, showing how GNN-based models can represent complex probability distributions associated with causal variables.
Together, these developments provide a promising foundation for building data-driven causal models of marine socio-ecological systems. Marine environments are ideal candidates for GNN-based approaches. They contain complex networks of interaction, nonlinear dependencies and spatially structured relationships. GNNs can learn from large and heterogeneous datasets, combining ecological observations, sensor data, satellite imagery and socioeconomic information. When integrated with causal structures, they can reveal which relationships are merely correlational and which represent deeper causal mechanisms.
For marine ecosystem management, this has significant implications. As digital twins of ocean regions become more advanced, incorporating GNNs could support real-time causal inference. Managers could test policy options, evaluate risks and anticipate system-wide effects of interventions. For example, a GNN informed by causal structure might help predict how changes in fishing pressure would cascade through a food web, or how shifts in plankton productivity might affect fisheries and local economies. Because GNNs can handle high-dimensional data, they can capture feedbacks that escape simpler models.
The EcoTwin report provides a clear introduction to how Graph Neural Networks can contribute to causal discovery in marine socio-ecological systems. By aligning GNN architectures with causal graphs, researchers can combine the strengths of deep learning with the interpretability and rigour of causal reasoning. As marine science moves toward increasingly integrated and data-rich modelling approaches, GNNs have the potential to play a central role in revealing patterns, supporting interventions and powering digital twins of the ocean.