Leveraging physics-informed neural networks for efficient modelling of coastal ecosystems dynamics: A case study of Sundarbans mangrove forest
Modelling the hydro-morphodynamics of mangrove environments poses significant challenges due to their complex geometry, dynamic vegetation-water interactions, and limited in situ measurements. Traditional numerical models struggle in such settings due to high computational cost, the need for extensi...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Ecological Informatics |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1574954125003115 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Modelling the hydro-morphodynamics of mangrove environments poses significant challenges due to their complex geometry, dynamic vegetation-water interactions, and limited in situ measurements. Traditional numerical models struggle in such settings due to high computational cost, the need for extensive mesh generation, and difficulties in assimilating heterogeneous data sources. To overcome these limitations, this study proposes a hybrid Physics-Informed Neural Networks (PINNs) framework that integrates governing physical laws, satellite-derived vegetation data, and outputs from high- resolution Computational Fluid Dynamics (CFD) simulations of the Sundarbans mangrove system. The proposed hybrid PINN model incorporates shallow water flow and sediment transport equations into the loss function and uses a temporal causality weighting mechanism to improve convergence in time-dependent domains. Model outputs include elevation, flow velocity, and suspended sediment concentration. The model is trained using a small subset of CFD simulation data and validated against a traditional finite element (FE) solver and a Convolutional Neural Network (CNN) baseline. The proposed hybrid PINNs model achieved RMSE values in the range of 10−2 for elevation and velocity and 10−3 for sediment concentration, significantly outperforming the CNN model, particularly in generalising to unseen time steps. Compared to the FE model, the proposed framework achieved a five-fold reduction in training time (24 h vs. 5 days) and supports inference within seconds, enabling near real-time predictions. The mesh-free structure of the model, combined with its physics-regularised learning, also facilitates adaptation to other coastal or mangrove systems governed by similar dynamics. This hybrid approach offers a robust, generalisable surrogate for high-fidelity coastal simulations in data-scarce and computationally demanding environments. |
|---|---|
| ISSN: | 1574-9541 |