Disentangling Internal and Forced Climate Variability with Convolutional Neural Networks using Multivariate Fields
Abstract
Long-term climate data exhibit variations composed of internal and forced components. Internal variability arises from natural processes that could occur within a stable climate. Forced variability, on the other hand, reflects climate changes induced, for example, by anthropogenic greenhouse gas and aerosol emissions. Accurately distinguishing between these types of variability is crucial for attributing climate fluctuations and understanding internal variability processes and impacts. In this study, we apply a U-Net convolutional network, a model commonly used in computer vision, to separate forced and internal climate variability from 1950 to 2022 using a multi-model dataset. The dataset includes multiple fields from five single-model initial-condition large ensembles. Cross-validation is conducted by training the U-Net using the data from the ensembles of four models, leaving out the data from one model to assess performance. Validation results yield errors ranging from 0.1°C to 0.4°C for the forced variability of local-scale monthly surface air temperature, which is no more than half the magnitude of external forcing. The U-Net achieves better performance than a simple approach based on a fourth-order polynomial trend for estimating forced variability. The error is mainly due to insufficient sampling and poor agreement among the models, as the U-Net underestimates the warming for the model with the highest transient climate sensitivity during validation. Performance is lower for variables like sea-level pressure and precipitation because of their low ratio of forced to internal variability. This framework might be enhanced by incorporating a larger multi-model dataset in the training and validation.
Related
- Disentangling forced and internal components in observed climate evolution using spatiotemporal deep learning
- Convolutional neural networks for enhanced classification of VLF rock rupture precursors
- Downscaling with AI reveals the large role of internal variability in fine-scale projections of climate extremes
- Reservoir eutrophication prediction with explainable framework integrating temporal convolutional networks and bidirectional sequence modeling
- Application of long short-term memory and convolutional neural networks for real-time bridge pier scour prediction
BibTeX
@misc{gastineau2026disentangling,
title = {{Disentangling Internal and Forced Climate Variability with Convolutional Neural Networks using Multivariate Fields}},
author = {Guillaume Gastineau and Elena Provenzano and Constantin Bône and Nicolas Lebas and Robert C. J. Wills},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.05359},
eprint = {2609.05359},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.05359},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Disentangling Internal and Forced Climate Variability with Convolutional Neural Networks using Multivariate Fields
AU - Guillaume Gastineau
AU - Elena Provenzano
AU - Constantin Bône
AU - Nicolas Lebas
AU - Robert C. J. Wills
PY - 2026
DA - 2026/09/04
PB - arXiv
DO - 10.48550/arxiv.2609.05359
UR - https://arxiv.org/abs/2609.05359
AB - Long-term climate data exhibit variations composed of internal and forced components. Internal variability arises from natural processes that could occur within a stable climate. Forced variability, on the other hand, reflects climate changes induced, for example, by anthropogenic greenhouse gas and aerosol emissions. Accurately distinguishing between these types of variability is crucial for attributing climate fluctuations and understanding internal variability processes and impacts. In this study, we apply a U-Net convolutional network, a model commonly used in computer vision, to separate forced and internal climate variability from 1950 to 2022 using a multi-model dataset. The dataset includes multiple fields from five single-model initial-condition large ensembles. Cross-validation is conducted by training the U-Net using the data from the ensembles of four models, leaving out the data from one model to assess performance. Validation results yield errors ranging from 0.1°C to 0.4°C for the forced variability of local-scale monthly surface air temperature, which is no more than half the magnitude of external forcing. The U-Net achieves better performance than a simple approach based on a fourth-order polynomial trend for estimating forced variability. The error is mainly due to insufficient sampling and poor agreement among the models, as the U-Net underestimates the warming for the model with the highest transient climate sensitivity during validation. Performance is lower for variables like sea-level pressure and precipitation because of their low ratio of forced to internal variability. This framework might be enhanced by incorporating a larger multi-model dataset in the training and validation.
ER -