midatm1234/prithvi-wxc-downscaling

Analyze Prithvi WxC downscaling inputs and outputs (preprocessed training NetCDF or inference output) through the prithvi-wxc-downscaling MCP tools — load a date, summarize variables, regional statistics, maps, date differences, trends, monthly/seasonal climatologies. Use when the user asks about values, plots, statistics, or climatology of downscaled precipitation, Tmax, Tmin, or the predictors.

Download the inputs for Prithvi WxC downscaling — MERRA-2 or NARR predictors, PRISM 800 m targets (ppt/tmax/tmin), static 800 m orography, pretrained weights, and optionally CORDEX-ML-Bench — via the prithvi-wxc-downscaling MCP download jobs. Use when the user asks to download, fetch, or check MERRA-2, NARR, PRISM, elevation/orography, weights, or training/inference data for a date range.

Train and run Prithvi WxC / Prithvi-UNet downscaling (32 km NARR or MERRA-2 to 800 m PRISM precipitation, Tmax, Tmin) through the prithvi-wxc-downscaling MCP tools — build a localized YAML config, preprocess, compute scalars, fine-tune, run tiled inference, and evaluate against PRISM as tracked background jobs. Use when the user asks to downscale, train, fine-tune, run inference, or change a downscaling config (region, dates, variables, epochs, GPUs).

Add stochastic residual refinement to NARR Prithvi-UNet downscaling — a diffusion or flow-matching model of the residual y_true - y_hat trained on top of the frozen deterministic Prithvi model, giving a bias-corrected ensemble (mean and spread) instead of one field. Use when the user asks for diffusion, flow matching, generative or stochastic refinement, residual correction, ensembles, or uncertainty for NARR downscaling. NARR only; MERRA-2 has no refinement.

Reproduce or share a Prithvi WxC downscaling run using run manifests — inspect what code commit, config, pins, interpreter, and inputs a job used; check whether it is reproducible; replay a manifest received from another machine or institution. Use when the user asks to reproduce, replay, re-run, compare, audit, or share a run, or asks "which version / config produced this".

Prepare a machine to run Prithvi WxC downscaling — check GPUs, disk, credentials, clone the pinned code, and build the pinned training environment through the prithvi-wxc-downscaling MCP tools. Use when the user is on a new machine, asks to "set up", "install", or "get started" with Prithvi WxC / Prithvi-UNet downscaling, or when another prithvi skill hits a missing-environment error.