Geophysical diversity can structure freshwater biodiversity by shaping habitat heterogeneity through hydrology, geomorphology, and thermal regimes. Conservation efforts to preserve biodiversity might prioritize preserving geophysical diversity and connectivity of habitats, operating under the assumption that habitat diversity within topological networks necessarily begets aquatic biodiversity. However, this hypothesis has not been adequately tested at a variety of spatial scales, levels of community organization ( , , and diversity), and transferability amongst divergent ecoregions. yet its predictive value is often unclear when models are transferred across broad ecological regions. We modeled reach-scale fish species richness based on testing a series of alternative hypotheses about what structures biodiversity: 1) universal predictors of biodiversity, such as Ecoregion, system size (i.e., discharge, drainage area), and terrestrial productivity (NPP), 2) geophysical habitat types (e.g., hydrological types, stream thermal regimes), and 3) diversity in geophysical habitat types. Alternative models were constructed using a Bayesian negative binomial generalized linear mixed models (NB-GLMM) with random intercepts for freshwater ecoregions.. We evaluate predictive skill both within ecoregions and under transfer to entirely new ecoregions, revealing strong within-region performance but markedly reduced transferability, consistent with region-specific structure and/or missing transferable gradients. Building on this baseline, we are extending analyses to quantify habitat diversity explicitly at multiple spatial scales (local “alpha” vs watershed “gamma”; e.g., HUC8 summaries) and to compare hierarchical models with machine-learning approaches that can capture nonlinearities and interactions while retaining interpretability. More flexible machine-learning (ML) models (e.g., gradient-boosted trees, random forests, or neural networks) are currently tested that can complement NB-GLMMs by capturing nonlinearities, high-order interactions, and thresholds in geodiversity–biodiversity relationships, while modern interpretation tools (e.g., SHAP/partial dependence) can retain ecological insight. Comparing ML and hierarchical statistical models under the same “new-ecoregion” validation can also identify which predictor sets and model structures generalize best, supporting transferable biodiversity mapping and conservation prioritization.