Predicting where multiple species occur is a major challenge when observations are sparse. Standard species distribution models treat each species independently and cannot capture the metacommunity dynamics that determine which species occupy which locations. We introduce MetaDiffusion, a conditional diffusion model trained on individual-based model (IBM) simulations to produce uncertainty-quantified predictions of multi-species community distributions. Training used 240 IBM simulations of competitive metacommunity dynamics across diverse parameters, producing per-species biomass maps and presence/absence time series on 20x20 grids. The model conditions on four inputs: local environmental suitability (CNN), inter-patch dispersal structure (GNN), temporal community dynamics (Transformer), and sparse field observations, combined through Feature-wise Linear Modulation into a U-Net denoiser, producing distribution ensembles with explicit uncertainty. Ablation experiments confirmed that temporal assembly history is the critical input: removing the competition graph reduced richness correlation by only r=0.004, whereas removing temporal history caused a collapse from r=0.938 to r=-0.024. Replacing temporal history with 5 to 10 sparse field observations per species, matching data availability for rare species in conservation records, maintained richness correlation at r=0.857 to r=0.914. Since 5 observations cover less than 2% of species-cell combinations, this recovery demonstrates the model uses learned ecological constraints rather than the observations directly. Validated on held-out IBM simulations, MetaDiffusion achieved AUC=0.853 for rare species, nearly matching common species. Richness correlation reached r=0.938, range-size correlation r=0.987, and beta-diversity correlation r=0.992, with range-size distributions statistically indistinguishable from IBM outputs (KS p=0.17). MetaDiffusion represents a mechanistically grounded, uncertainty-aware alternative to correlative species distribution models for biodiversity assessment. This abstract was prepared for submission to the European Conference on Computational Biology (ECCB) 2026. A full manuscript is in preparation.
Shahin et al. (Tue,) studied this question.