AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

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

Minseo Yoon, Minseong Bae, Sojin Lee, Hyunwoo J. Kim

NeurIPS 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: Low-rank compression and merging · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PROLA, a principal-orthogonal low-rank adaptation method for predictive spatiotemporal weather downscaling (PSWD), producing future high-resolution weather fields from recent coarse states.

Problem

Adapting large pretrained weather backbones to predict future high-resolution fields from recent coarse atmospheric states is challenging and costly.

Approach

Formulates predictive spatiotemporal weather downscaling as a task and adapts pretrained weather backbones via principal-orthogonal low-rank adaptation (details beyond the abstract not available).

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use PROLA dual-branch adaptation instead of LoRA at rank 8Aurora PSWD · WeatherBench 2 3h Z500RMSE: 54.72 -> 50.58Improved
use PROLA instead of PiSSAAurora PSWD · ERA5-to-CERRA 3h Z500RMSE: 71.80 -> 68.44Improved
use PROLA instead of OPLoRAAurora · Regional precipitation 12hSEEPS: 0.605 -> 0.590Improved
PROLA (3.92M params) vs full fine-tuning (1.26B)Aurora PSWD · WeatherBench 2 Z500RMSE: 60.13 -> 50.58 (3h)Improved
add spatial and temporal correction heads (ablation)PROLA pipeline · WeatherBench 2 3h Z500RMSE: 87.73 -> 50.30Improved
add temporal correction head Ct (ablation)PROLA pipeline · WeatherBench 2 3h Z500RMSE: 84.53 -> 50.30 (vs w/o Ct)Improved
add spatial correction head Cs (ablation)PROLA pipeline · WeatherBench 2 T2M/U10 6hRMSE: 6h Z500 78.25 -> 73.07; T2M 1.209 -> 1.211Mixed
use scaffold-and-correct pipeline instead of direct prediction (ablation)PROLA · WeatherBench 2 3h Z500RMSE: 53.13 -> 50.30Improved
add orthogonal branch to principal branch (4+4) (ablation)PROLA · WeatherBench 2 3h Z500RMSE: 53.66 -> 51.36Improved
add gradient-guided update scaling (ablation)PROLA 4+4 · WeatherBench 2 3h Z500RMSE: 51.36 -> 50.30Improved
use balanced 4+4 rank split instead of 6+2 or 2+6 (ablation)PROLA · WeatherBench 2 6h Z500RMSE: 73.07 vs 74.19 and 76.35Improved
use both branch and component scores instead of one (ablation)PROLA · WeatherBench 2 6h Z500RMSE: 76.06 / 75.45 -> 73.07Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Experimental setup

Builds on: LoRA

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy