Topological Graph Generative Model for Ecological Design

Designing ecological systems, whether for conservation planning, microbial communities, or tissue architectures, requires anticipating how populations will evolve. Yet most existing tools optimize indirect measures that do not reliably predict long-term evolutionary outcomes, limiting their usefulness across domains. We introduce a new approach that treats ecological design as a graph generation problem grounded in evolutionary theory. In this framework, fragmented ecosystems or connectivity networks of interacting individuals in the population are represented as graphs, and we guide their design using two intuitive controls, based on prior theoretical work: amplification, the likelihood that a new trait will spread, and acceleration, the speed at which this happens. To train our model, we create a dataset of 12,173 synthetic networks, each evaluated through evolutionary simulations. We then develop a generative model that proposes designs and a “compile-to-edits” step that translates them into practical, budgeted modifications, while ensuring the network remains connected. Our method exhibits calibrated target to realized control for both factors, uncovers clear structural patterns, produces fixation curves consistent with targets across selection strengths, and outperforms baselines when applied to a real conservation case in the Eldorado National Forest. This work delivers the first end-to-end, evolution-aware generative design tool, advancing principled and budget-conscious ecological interventions.