AbstractEcological simulations including forest and vegetation growth models require climate inputs that match the resolution and extent of the process being modelled. Climate inputs are often derived at resolutions coarser than the scale of many ecosystem processes. Machine learning models can be trained to spatially downscale climate data to fine (30 m) resolution using topographic variables such as elevation, aspect and other site‐specific factors. Statistically downscaled climate models will have spatially varying uncertainty that is not usually incorporated into downscaling techniques for error propagation into later models, are often applied on smaller areas, are not fine enough resolutions for many modelling techniques, or are not always scalable to large spatial extents. There remains opportunity to leverage machine learning advancements to downscale climate to very fine (30 m) resolutions with associated spatially explicit uncertainty to represent microclimatic variation in ecological models. In this study, we used quantile machine learning to produce 30 m downscaled temperature and precipitation data and associated model prediction uncertainty for the state of California. Temperature models were accurate at downscaling 4 km climate data to 30 m, performing better than the 4 km data at high and low slope positions and at high elevations, especially where there were fewer weather observations. Precipitation model predictions did not show global improvement over the 4 km scale, but were more accurate at high elevations, slopes with higher solar radiation and in valleys. For all climate variables, the added detail of spatial explicit uncertainty via 90% prediction intervals provides critical insight into the utility of empirically downscaled climate. The resulting 30 m spatially contiguous outputs can be used as ecological model inputs with uncertainty propagation, to illuminate climate trends over time as a function of fine‐scale spatial factors, and to highlight areas of spatially explicit uncertainty.