Transferability of ML emulators for polar RCM surface melt under different forcings: Limitations and recommendations
Abstract
Machine learning emulators provide a computationally efficient alternative to physically based regional climate and surface mass balance models, but their applicability depends on their ability to generalize beyond training conditions. In this study, we assess the transferability of a melt emulator trained on output from the regional climate model HIRHAM5 forced by ERA-Interim, CESM2, and EC-Earth under historical and future (SSP5-8.5) climates. Transferability is evaluated across forcing datasets and climate states without retraining. The emulator generalizes reasonably well within a given climate regime, producing biases below the inter-model spread, but exhibits limited transferability between historical and future climate conditions. Predictor distributions analyses, perturbation experiments, and ablation studies indicate that these errors arise from shifts in predictor distributions together with an insufficient representation of the surface state. Although perturbation experiments demonstrate physically plausible relationships between surface energy fluxes and melt generation, physical consistency alone does not ensure robust extrapolation to novel climate conditions. Increasing training data diversity does not overcome these transferability limitations, highlighting the importance of training on data that are representative of the target application. Future emulator development should focus on incorporating information across multiple temporal scales and systematically evaluate transfer performance to improve generalizability.
Related
- Time-mean Tropical Circulation in an Idealized Moist GCM: Dependence on Zonally Asymmetric Forcings
- Influence of natural and anthropogenic forcings on western US winter fire weather and the 2025 Los Angeles wildfires
- Machine learning emulators of dynamic global vegetation models for fast gridded carbon budget predictions
- SamudrACE: Fast and Accurate Coupled Climate Modeling With 3D Ocean and Atmosphere Emulators
- Towards accurate extreme event likelihoods from diffusion model climate emulators
BibTeX
@misc{schlager2026transferability,
title = {{Transferability of ML emulators for polar RCM surface melt under different forcings: Limitations and recommendations}},
author = {Elke Schlager and Sebastian Scher and Ruth Helen Mottram and Peter Langen},
howpublished = {EarthArXiv},
year = {2026},
month = sep,
doi = {10.31223/x5cr59},
url = {https://eartharxiv.org/repository/view/15294/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Transferability of ML emulators for polar RCM surface melt under different forcings: Limitations and recommendations
AU - Elke Schlager
AU - Sebastian Scher
AU - Ruth Helen Mottram
AU - Peter Langen
PY - 2026
DA - 2026/09/30
PB - EarthArXiv
DO - 10.31223/x5cr59
UR - https://eartharxiv.org/repository/view/15294/
AB - Machine learning emulators provide a computationally efficient alternative to physically based regional climate and surface mass balance models, but their applicability depends on their ability to generalize beyond training conditions. In this study, we assess the transferability of a melt emulator trained on output from the regional climate model HIRHAM5 forced by ERA-Interim, CESM2, and EC-Earth under historical and future (SSP5-8.5) climates. Transferability is evaluated across forcing datasets and climate states without retraining. The emulator generalizes reasonably well within a given climate regime, producing biases below the inter-model spread, but exhibits limited transferability between historical and future climate conditions. Predictor distributions analyses, perturbation experiments, and ablation studies indicate that these errors arise from shifts in predictor distributions together with an insufficient representation of the surface state. Although perturbation experiments demonstrate physically plausible relationships between surface energy fluxes and melt generation, physical consistency alone does not ensure robust extrapolation to novel climate conditions. Increasing training data diversity does not overcome these transferability limitations, highlighting the importance of training on data that are representative of the target application. Future emulator development should focus on incorporating information across multiple temporal scales and systematically evaluate transfer performance to improve generalizability.
ER -