Weather on the Edge: Client-Side Post-Processing For Mobile Weather Applications

Delivering hyper-local weather forecasts on mobile typically places a heavy burden on backend infrastructure: point-based post-processing services must either pre-compute forecasts for large sets of locations or respond to on-demand requests at scale. Both approaches carry growing infrastructure costs and require users to share their location with a server. We propose an alternative architecture that inverts this model: a lean backend paired with an intelligent client that runs post-processing entirely on device.

European weather data is anything but uniform: providers publish in different file formats or APIs, on different schedules, and in different map projections. Our backend ingests data from 15 services and normalizes them into a single, unified archive, stored in a cloud-optimized Zarr format for fast timeseries access at scale. Map products are pre-rendered as compressed raster tiles.

Our post-processing pipeline runs on the edge: We leverage the power of modern mobile devices to combine dynamic weather data with a static, high-resolution digital elevation model (DEM). This approach has several benefits:

1. Small data transfers. Very high resolution weather map animations can be achieved through on-device upsampling and interpolation without requiring prohibitively large data transfers. For example, we demonstrate how temperature maps with kilometer-scale resolution can be upsampled in real-time by on-device GPU’s to create a more realistic, interactive temperature field with a resolution of tens of meters.

2. Distributed compute. Hyper-local weather forecasts are usually achieved through post-processing forecasts for a large set of pre-defined points of interest or running post-processing on demand as a point-based backend service. Running post-processing on device achieves the hyper-local quality while keeping infrastructure costs managable.

3. Privacy. Being fully aware of all data ingested into our Zarr archive, our client chooses which data to load and autonomously runs post-processing for their exact location without it ever leaving the device.

Together, these properties enable a class of weather application that was previously out of reach: one that is simultaneously hyper-local, privacy-preserving, and visually rich — without demanding heavy server-side compute. We present Fluid Meteo as a working implementation of this approach, and discuss the practical trade-offs encountered in building and operating it.