---
format: "aidr-story-markdown/v1"
id: "e1a659b2c72259899230e150131fb11bd91048d02b4e5eb289faa35e972029e2"
canonical_url: "https://aidr.today/e1a659b2?lang=en"
title: "PROLA: Principal-Orthogonal Low-rank Adaptation for Predictive Spatiotemporal Weather Downscaling"
lang: "en"
requested_lang: "en"
available_langs: ["en"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-19T00:00:00.000Z"
category: "Research"
topics: ["inference"]
source_urls: ["https://huggingface.co/papers/2611.11318"]
summary: "High-resolution weather prediction is important for resolving local atmospheric patterns missed by coarse global forecasts. Large pretrained weather backbones have improved global forecasting, but adapting them to produce future high-resolution fields from recent coarse states remains challenging and costly. We formulate this setting as Predictive Spatiotemporal Weather Downscaling (PSWD), a task that jointly refines spatial and temporal resolution while predicting future high-resolution trajectories. To this end, we propose a framework that decomposes prediction into spatial and temporal stages using a shared pretrained backbone, fixed numerical scaffolds, and lightweight residual heads. For backbone adaptation, we introduce PROLA (PRincipal-Orthogonal Low-rank Adaptation), which allocates a fixed low-rank budget between pretrained weights' principal directions and their orthogonal complement. PROLA further rescales rank-1 optimizer updates using gradient signals, improving adaptation without increasing trainable rank. Experiments in the PSWD setting show that PROLA outperforms representative low-rank adaptation baselines."
---

# PROLA: Principal\-Orthogonal Low\-rank Adaptation for Predictive Spatiotemporal Weather Downscaling

> [Open the canonical story](<https://aidr.today/e1a659b2?lang=en>)

**Published:** 2026-11-19T00:00:00.000Z
**Category:** Research
**Topics:** inference

## Summary

High\-resolution weather prediction is important for resolving local atmospheric patterns missed by coarse global forecasts\. Large pretrained weather backbones have improved global forecasting, but adapting them to produce future high\-resolution fields from recent coarse states remains challenging and costly\. We formulate this setting as Predictive Spatiotemporal Weather Downscaling \(PSWD\), a task that jointly refines spatial and temporal resolution while predicting future high\-resolution trajectories\. To this end, we propose a framework that decomposes prediction into spatial and temporal stages using a shared pretrained backbone, fixed numerical scaffolds, and lightweight residual heads\. For backbone adaptation, we introduce PROLA \(PRincipal\-Orthogonal Low\-rank Adaptation\), which allocates a fixed low\-rank budget between pretrained weights' principal directions and their orthogonal complement\. PROLA further rescales rank\-1 optimizer updates using gradient signals, improving adaptation without increasing trainable rank\. Experiments in the PSWD setting show that PROLA outperforms representative low\-rank adaptation baselines\.

## Sources

- [Story source](<https://huggingface.co/papers/2611.11318>)

