Ice sheets
Neural nets map CMIP6 weather into Greenland melt
Open access · cc by · source: Europe PMC
ANNs trained on CESM2 translate CMIP6 atmospheres into Greenland surface-melt increases of about 79–264% by 2100 depending on SSP.
How this study connects
Role on claims
Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.
ANNs trained on CESM2 summer T2m and snowfall, then applied to hundreds of CMIP6 historical and SSP runs, raise Greenland Ice Sheet melt 79% (SSP1-2.6) to 264% (SSP5-8.5) versus 1979–1998 (414±276 to 1,378±555 Gt yr⁻¹). SSP5-8.5 ensemble-mean bias versus MAR is only 1.6%, but SSP1-2.6 is 71.9% high. Networks inherit one GCM’s melt physics; this is surface melt, not full sea-level contribution.
Evidence for the claim as stated.
Emulators trained on one GCM can look excellent in one scenario and fail in another. The Greenland melt ANN’s SSP5-8.5 ensemble mean sits 1.6% from MAR, but SSP1-2.6 is 71.9% high; climate-sensitivity spread across CMIP6 dominates. Prescribed-ice CESM2 ensembles (Arctic sea ice −19.0% winter, −48.4% summer) are not downscaling at all: they are large-ensemble sensitivity experiments whose melt uncertainty (±64 Gt yr⁻¹ in summer SMB) dwarfs the winter snowfall gain (23±33 Gt yr⁻¹).
Evidence for the claim as stated.
Open questions
Tensions this paper is part of
From concept pages' “where studies disagree.” Disagreement means the same question; scope means different assays, populations, or outcomes.
Emulators trained on one GCM can look excellent in one scenario and fail in another. The Greenland melt ANN’s SSP5-8.5 ensemble mean sits 1.6% from MAR, but SSP1-2.6 is 71.9% high; climate-sensitivity spread across CMIP6 dominates. Prescribed-ice CESM2 ensembles (Arctic sea ice −19.0% winter, −48.4% summer) are not downscaling at all: they are large-ensemble sensitivity experiments whose melt uncertainty (±64 Gt yr⁻¹ in summer SMB) dwarfs the winter snowfall gain (23±33 Gt yr⁻¹).
Study at a glance
- Design
- Computational / modelling — ANNs trained on CESM2 melt then applied across CMIP6 SSP atmospheres
- N
- Hundreds of CMIP6 historical/SSP runs; ensemble size not a single primary N in stored text
- Population
- Greenland Ice Sheet surface melt under CMIP6 21st-century scenarios
- Outcome
- Projected melt increases (~79–264% by 2100 depending on SSP)
Structured fields used in claim comparison tables when every cited study has a complete layer.
Inspect source passages
Neural nets map CMIP6 weather into Greenland melt
These passages occur verbatim in the stored paper text. They support the points listed below; they do not verify every statement in the explanation.
How was training data split?
These CESM2 simulation years were randomly divided into two parts: the first to train the neural networks (90%), and the second for cross‐validation (that is, to tune the neural networks, 10%).
How many CMIP6 runs feed the projections?
In total, we use 345, 123, 147, 147, and 129 simulations from 46, 30, 32, 28, and 32 climate models for the historical period, SSP1‐2.6, SSP2‐4.5, SSP3‐7.0, and SSP5‐8.5, respectively.
Key findings
Melt rises 79% (SSP1-2.6) to 264% (SSP5-8.5) versus 1979–1998. Absolute increases are 414±276 to 1,378±555 Gt yr−1. T2m and snowfall ANNs beat cloud/Z500/radiation versus MAR. SSP5-8.5 ensemble mean bias versus MAR is only 1.6%.
Methodology
Authors trained separate ANNs on CESM2 summer T2m, Z500, cloud, radiation, and snowfall to predict annual GrIS melt, then applied T2m and snowfall networks to hundreds of CMIP6 historical and SSP runs and compared a subset with MAR.
Limitations
Networks are trained on one GCM’s melt physics; CMIP6 climate-sensitivity spread dominates; SSP1-2.6 MAR comparison is much worse (71.9% high); this is surface melt, not full sea-level contribution.
Related papers in this topic
Same topic cluster — not a recommendation engine.