Modelling the impact of linear infrastructure on species dispersal in a changing climate: a scalable surrogate approach

International audience

Inglés. As climate change accelerates, many species must shift their ranges to survive. Yet humanmade linear infrastructures-such as roads, railways, and power lines-fragment habitats and create major barriers to movement. Understanding and mitigating these effects is now an urgent task for ecologists, conservation planners and infrastructure maintenance managers. Individual-based models like RangeShifter offer a powerful way to simulate how populations respond to changing environments and landscape structures. They incorporate detailed mechanisms such as dispersal behavior, life history, and even natural selection. However, these models are computationally intensive, and their use becomes impractical when applied to large spatial scales or long-term climate scenarios-precisely the scales needed to inform real-world adaptation strategies.Our work addresses this challenge by developing a surrogate model-a computational shortcut-that mimics the behavior of RangeShifter but runs much faster. This is achieved through a two-step process: first, we optimize the original simulator to generate training data efficiently, through a OpenMP parallelization process on cluster. Then, we train a deep learning model, specifically a conditional denoising diffusion probabilistic model (DDPM), to reproduce RangeShifter's output of the population distribution conditionally on the landscape map, the population state at the previous timestep, and the species' ecological parameters. To make this training feasible, we embed the DDPM within a latent space learned by a pretrained variational autoencoder (VAE), allowing the model to learn from compressed representations of complex simulation outcomes. The result will be a surrogate model that can explore a vast space of landscape and species scenarios in a fraction of the time, enabling ecological forecasting at realistic scales. This tool is designed to support ecologists in evaluating how different infrastructure designs or mitigation strategies might influence connectivity and long-term population viability under climate change.