Plant traits are important for understanding community assembly and ecosystem processes, yet our understanding of intraspecific trait variation (ITV) is limited. This gap in our knowledge is partially because collecting trait data across a species’ entire range is impractical, let alone across the ranges of multiple species within a plant family. Using machine learning techniques to predict spatial ITV is an attractive and cost-effective alternative to sampling across a species range, although this has not been applied beyond regional scales. We compiled a trait database of over 1,000 grass species (family: Poaceae), encompassing six key functional traits: specific leaf area (SLA), leaf dry matter content (LDMC), plant height, leaf area, leaf nitrogen (Nmass) and leaf phosphorus content (Pmass). Using a random forest machine learning approach, we predicted local trait values within species' ranges considering climate, soil type, phylogeny, lifespan, and photosynthetic pathway as influential factors. An iterative random forest modeling technique incorporated correlations between traits, resulting in improved model performance (observed vs. predicted R2 range of 0.72 - 0.91). Our models also highlight the importance of climate in predicting trait variation. For a subset of species (n = 860), we projected trait predictions across their known distribution, informed by expert maps from Kew Botanical Gardens, to create global maps of ITV for grasses. Such maps have the potential to inform conservation efforts and predictions of grazing and fire dynamics in grasslands worldwide. Overall, our research demonstrates the value and ecological applications of predicting plant traits.