Global climate models are too coarse to describe a particular place. I tested two ways of sharpening them, both in the cloud.
Coarse climate dataGlobal model output
→
Fine-tuned foundation modelIBM–NASA Prithvi WxC
→
Fine-resolution estimateTemperature and wind
→
EvaluationCompared with ClimaX
CMIP6 projectionsCoarse global models
→
Sup3rCC pipelineReproduction of NREL's sup3r downscaling models
→
Fine-resolution estimateReproducible, toward production
The two approaches I evaluated. Architecture sketch.
My role
Developer, working with climate scientists
Years
2025
Built with
AWS SageMaker, PyTorch, Python
Worked with
ICF's ClimateSight team
The problem
Climate risk work needs local answers, but global models describe the climate in large grid cells. Downscaling bridges the gap.
What I did
Fine-tuned the IBM–NASA Prithvi WxC foundation model to downscale two-metre temperature from MERRA-2 and surface winds from ECCC data, and compared it with ClimaX.
Reproduced NREL's sup3r approach to downscaling CMIP6 projections, and built the pipeline so it can be re-run and moved toward production.
Ran the training and evaluation on AWS SageMaker, so the work did not depend on a single machine.
Why it matters
A foundation model that can be adapted to a new region or variable with modest data is a different kind of tool from a bespoke model per project. This work tested whether that holds up for downscaling.
Outcome
A working comparison of the two approaches and a reproducible pipeline, delivered to ICF's ClimateSight team.