Benchmarks, datasets & evaluation
AI-Based Climate Change Forecasting Models
In one sentence. Compares six AI approaches across four climate tasks, claiming hybrid physics-ML best at 6-month temperature RMSE; claims are unverified and methodology thin.
Abstract
Climate change forecasting demands models that capture complex nonlinear interactions across atmospheric, oceanic,and terrestrial systems at spatial and temporal scales ranging from regional weather patterns to century-scale globaltrends. Physics-based general circulation models (GCMs) remain the gold standard for climate projection but arecomputationally expensive and struggle with sub-grid-scale processes that must be parameterised rather than resolved.This study evaluates six AI-based approaches -- convolutional LSTMs, graph neural networks on climate grids, Fourierneural operators (FNO), vision transformers applied to reanalysis fields, hybrid physics-ML models (ML-correctedGCMs), and ensemble gradient-boosted models for regional downscaling -- across four climate forecasting tasks: globaltemperature anomaly prediction (ERA5, 1979-2021, 1-12 month horizons), regional precipitation forecasting (E-OBSEuropean gridded data, seasonal), extreme event detection (tropical cyclone genesis from ERA5 fields), and statisticaldownscaling (CMIP6 coarse-resolution to 0.1-degree regional fields). A total of 2,160 experiments were conducted usingstandardised train-validation-test splits respecting temporal ordering. The hybrid physics-ML model achieved the lowesttemperature RMSE at 6-month horizon (0.24 +- 0.03 K vs. 0.38 +- 0.04 K for the persistence baseline and 0.32 +- 0.04 Kfor the pure ML ConvLSTM), confirming that physics-informed architectures outperform pure data-driven approaches onclimate time scales. The FNO achieved the best spatial pattern reconstruction (SSIM = 0.92 +- 0.02 for globaltemperature fields) at 1,000x lower computational cost than equivalent GCM resolution. For extreme events, the visiontransformer achieved the highest tropical cyclone genesis detection AUC (0.924 +- 0.008) by capturing large-scaleatmospheric patterns across hemispheric fields. Gradient-boosted downscaling achieved the smallest regionalprecipitation bias (2.4 +- 0.8 mm/month) at 50x lower cost than dynamical downscaling. A practical model selectionframework mapping forecast horizon, spatial resolution requirements, and computational budget to recommended AIapproaches is proposed.
Labels
- Topic
- Benchmarks, datasets & evaluation, Sub-seasonal to seasonal, Downscaling & super-resolution
- Type
- Evaluation or intercomparison of existing models
- Method
- CNN / U-Net, Graph neural network, Neural operator (FNO, SFNO, ...), Transformer / ViT, Physics-informed / physics-constrained, Classical ML (trees, GP, linear, kernels)
- Time scale
- Seasonal to annual
- Domain
- Global, Regional
- Data
- ERA5 / reanalysis, CMIP / climate model output, Other
- Status
- Research
- Identifiers
- DOI:10.5281/zenodo.19614573 OpenAlex:W7203708720
- Added
- 2026-10-22
Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.