Arun0808/pc-rf-climate-downscaling

Physics-Constrained Rectified Flow (PC-RF) — Climate Downscaling Demo

A conditional rectified-flow model (PyTorch, exported to ONNX) that downscales coarse climate fields while respecting three atmospheric conservation laws: divergence-free wind, non-negative precipitation, and domain mass balance. The inference-time projection drives the non-negativity violation rate (NVR) to 0 and the mass-conservation error (MCE) to ~0.

This Space runs the exported ONNX velocity field (2.4 MB) on a free CPU: it samples the rectified-flow ODE, applies the physics projection, and reports the three physics-validity metrics (DE, NVR, MCE). The example fields are synthetic (generated by the paper's pipeline); the method targets ERA5 and Sentinel-2.

Hardware: CPU Basic (free). The model is sub-second on CPU, so no GPU is needed.

Companion to the AIAS 2026 paper and the PC-RF platform (training, ONNX / AOTInductor export, Rust serving, Temporal orchestration, vLLM agent).