CGIAR Climate Action CMIP7 downscaling — methods and infrastructure
What this page covers
The methods, compute, and data-sharing infrastructure needed to produce a CGIAR Climate Action African-tuned CMIP7 downscaled and bias-corrected dataset — the data that will eventually replace NEX-GDDP-CMIP6 as the operational backbone of the AAA Adaptation Atlas and the Build a Climate Rationale notebook.
This is the upstream half of CGIAR Climate Action’s CMIP7 production effort. The downstream half — the hazards pipeline that turns these downscaled inputs into adaptation-decision indicators — is the companion page.
For the partner-facing transition overview (what CMIP7 is, when it lands, what proposal authors should do today), see What’s next — CMIP7 and CORDEX-Africa.
Why CGIAR is in this business
No CMIP7-derived downscaled product for African adaptation exists yet from any team — NEX-GDDP-CMIP7 has no public release timeline, no African regional centre (ICPAC, AGRHYMET, ACMAD, CSAG) has announced a CMIP7-tuned product of its own, and the AAA Adaptation Atlas’s single largest dependency (NEX-GDDP-CMIP6) is at uncertain funding continuity through the FY2026–FY2027 US budget cycle (leadership brief).
A CGIAR-led African-tuned CMIP7 downscaling pipeline is both an opportunity and a hedge — opportunity to claim production-space for African adaptation, hedge against external dependency failure. This page is where that pipeline is documented.
Proposition — scope
To be drafted by Carlos Navarros. Suggested points to cover:
- Continental scope — sub-Saharan Africa + Madagascar; with consideration of North Africa scope.
- Variables — minimally daily tasmin / tasmax / pr to match NEX-GDDP-CMIP6 v2 baseline; aspirationally the full 9-variable NEX-GDDP variable list to support direct substitutability.
- Spatial resolution target — 0.25° (NEX-GDDP parity) vs higher (CHELSA-class) trade-off.
- Temporal coverage — full CMIP7 AFT horizon (1850–2100 historical+scenario).
- Scenarios — the seven CMIP7 scenarios (H, HL, M, ML, L, LN, VL) per van Vuuren et al. 2026.
- Model count target — the AFR-13 / AFR-8 / FULL-18 sub-ensemble logic from African CMIP6 Ensembling, re-derived against CMIP7 model identities.
Bias-correction methodology
To be drafted by Carlos Navarros. Sources for the existing literature:
- BCSD (Bias Correction / Spatial Disaggregation) — the NEX-GDDP-CMIP6 method (Thrasher et al. 2022, Scientific Data).
- QDM (Quantile Delta Mapping) + QPLAD — Climate Impact Lab’s GDPCIR method (Gergel et al. 2024, GMD).
- ISIMIP3BASD (trend-preserving bias-adjustment) — the ISIMIP method (Lange 2019, GMD).
- Delta-change methods — CHELSA, WorldClim approach.
Suggested decision points:
- Which method best handles African precipitation extremes (relevant to East African paradox handling)?
- How to handle the observational baseline — CHIRTS / CHIRPS (Africa-tuned, the existing AAA convention) vs ERA5-Land (global, longer record) vs MSWEP?
- Reference period for bias correction — 1995–2014 (CMIP6 standard) or shift to a longer window?
Downscaling methodology
To be drafted by Carlos Navarros.
- Statistical vs dynamical — for the operational pipeline, statistical (BCSD / QDM-class) is the right choice for breadth and speed; dynamical (CORDEX-Africa-CMIP7) is a complementary parallel product, not a primary.
- Spatial interpolation method — bilinear vs bicubic vs constructed-analogue.
- Topography handling — Ethiopian highlands, Atlas Mountains, East African Rift require careful treatment.
Compute infrastructure
To be drafted by Carlos Navarros. The infrastructure decision must support both this downscaling pipeline AND the downstream hazards pipeline — shared compute is one of the design constraints.
Suggested points to cover:
- Compute envelope — TB of input data ingested, TB of output, CPU-hour budget for the full ensemble × scenario × year matrix.
- Hosting options — AWS (NASA NEX precedent), Google Cloud, Microsoft Azure (Planetary Computer precedent), Copernicus DIAS (EU), CGIAR-internal HPC (Alliance Bioversity-CIAT, ILRI, others).
- Decision drivers — cost, data residency, partner accessibility, durability.
- Workflow orchestration — Pangeo / Dask, Snakemake, Prefect, Argo Workflows, others.
Data storage and sharing infrastructure
To be drafted by Carlos Navarros.
Suggested points to cover:
- Output format — netCDF-CF, Zarr, GeoTIFF, COG, Parquet (for tabular indices).
- Access pattern — bulk download via S3 / Cloud Storage, OPeNDAP, STAC catalogue.
- Catalog / metadata — ESGF integration vs standalone catalogue vs Climate Data Hub native registry.
- License — CC-BY-4.0 (matching CMIP7 forcings convention).
- DOI and versioning — Zenodo, datacite, monthly or quarterly minted versions.
- Discoverability — partner-facing how-to-access guide; OGC API; STAC.
Pipeline architecture
To be drafted by Carlos Navarros.
Suggested stages:
- Ingest — pull raw CMIP7 model output from ESGF-NG / Copernicus C3S CDS, cache per model × scenario × variable.
- Bias-correct — apply the chosen method against the African observational baseline.
- Downscale — apply the chosen spatial method to the target grid.
- Validate — compare downscaled output against held-out observational sets; flag anomalies.
- Publish — emit netCDF/Zarr to storage, mint catalogue entry, update STAC.
- Hand off — make outputs available to the hazards pipeline.
Timeline and dependencies
To be drafted by Carlos Navarros.
| Quarter | Milestone | Dependency |
|---|---|---|
| 2026 Q4 | Funding decision; team scoping | CGIAR Climate Action leadership |
| 2027 Q1 | Pilot pipeline against 2–3 CMIP7 models | First AFT models on ESGF-NG |
| 2027 H2 | Scale to full AFR-13 successor ensemble | Continued ESGF-NG publication |
| 2028 H1 | First public release of CGIAR-AAA-CMIP7 v0.1 | Validation against observational baselines |
| 2028 H2 / 2029 | Integration into the AAA Adaptation Atlas Future Projections layer | Hand-off to the hazards pipeline complete |
Validation and QA/QC
To be drafted by Carlos Navarros.
- Reference data — CHIRTS / CHIRPS for Africa observational backbone; ERA5-Land as cross-check.
- Validation metrics — KGE, NSE, percent-bias for the historical period; trend preservation across the bias correction.
- Comparison products — NEX-GDDP-CMIP6 v2 (regression test), CIL-GDPCIR-CMIP6, CHELSA-CMIP6.
- AR6-region-resolved validation for the AR6 reference regions over Africa (Iturbide et al. 2020).
- Adversarial cases — known-difficult regions (East African long rains, Sahel summer monsoon, Madagascar orographic precipitation).
Open design questions
To be answered as part of the pilot phase:
- Single CGIAR-AAA-CMIP7 product or a small family of products (e.g. one for daily indices, one for bioclim variables, one for extremes)?
- Centralised CGIAR pipeline or partnership-led (CSAG, ICPAC contributing methodology)?
- How to integrate with ISIMIP4 and CORDEX-Africa-CMIP7 outputs rather than duplicate them?
- Maintenance commitment — how many years post-release will CGIAR support this product?
Further reading
- What’s next — CMIP7 and CORDEX-Africa — the partner-facing transition overview.
- CMIP7 hazards pipeline — the companion downstream page.
- African CMIP6 Ensembling — the existing CMIP6 methodology this pipeline will replace.
- Dataset landscape (draft) — the wider CMIP-era dataset catalogue.
- research_anchors_cmip7.md §10 — the full ecosystem map of parallel downscaling teams.
- CMIP7 leadership brief — the case for CGIAR claiming production-space.