Probabilistic Emulation of a Global Climate Model with Spherical DYffusion
This paper discusses a new model that helps simulate climate changes over long periods using advanced machine learning techniques. It aims to make climate predictions more accurate and less resource-intensive.
Content & Liability Disclaimer
This article and its accompanying video are automated summaries derived from the original research paper by Unknown authors. The original research was conducted solely by the paper's authors; PDFdigest did not conduct any of the research and makes no claims of ownership over the underlying scientific work.
The video narration is generated by artificial intelligence and references the paper's authors for attribution. The video is not narrated by any of the paper's authors. This content may contain inaccuracies, omissions, or misinterpretations of the original research. First-person language (e.g., "we found", "our results") reflects the original authors' voice, not PDFdigest's. Always read the original paper for accurate, verified information before making any decisions based on this content.
This content is provided "as is" without any warranties, express or implied. Simulated systems OÜ, its officers, directors, employees, and agents shall not be liable for any direct, indirect, incidental, special, consequential, or punitive damages arising from your use of, reliance on, or access to this content, including but not limited to errors, omissions, or misinterpretations of the original research. This disclaimer applies to the fullest extent permitted by applicable law.
- 1 ACE was trained on a next-step forecasting objective based on a MSE loss.
- 2 ACE is trained to emulate the United States' primary operational global forecast model, the physics-based FV3GFS , which is operationally used at the US National Weather Service and US National Centers for Environmental Prediction.
- 3 While it is possible to ensemble a deterministic model by perturbing its inputs, this approach often leads to under-dispersed (i.e. overly confident) ensembles compared to generative or physics-based approaches .
- 4 We show that short-term weather performance does not necessarily translate to accurate reproduction of long-term climate statistics.
Introduction
Such climate simulations are currently very expensive to generate due to the computational complexity of the underlying physics-based climate models, which must be run on supercomputers. Although recent deep learning models are on the verge of transforming the conceptually similar field of medium-range weather forecasting , these advances do not directly transfer to long-term climate projections .
We have verified that this finding holds regardless of the analyzed variable and the proxy used for weather performance, which we discuss in more detail in Appendix E.3.
While this is a little-discussed observation in the ML community, the climate modeling community has documented it for physics-based models .
SFNO and ACE are deterministic models that cannot be readily used for uncertainty quantification or ensemble-based climate modeling.
In our paper, these forcings correspond to prescribed sea surface temperatures and incoming solar radiation (see Section 5.1), leaving it to future work to force based on greenhouse gas emission scenarios explicitly.
Research Question
ACE was trained on a next-step forecasting objective based on a MSE loss.
Methodology
Unfortunately, the original DYffusion method relies on an UNet-based architecture designed for Euclidean data rather than physical fields on a sphere. For critical fields, such as the derived total water path quantity, our method achieves results within 20% of the reference model, representing a 5\u00d7 improvement over the next best baseline (see Fig. 2 ).
Study Design
Additionally, our method proves effective for ensemble climate simulations, reproducing climate variability consistent with the reference model and further reducing climate biases towards the theoretical minimum through ensemble-averaging.
The subsequent bars show the corresponding scores for our method and the deep-learning baselines, using a 25-member ensemble for the probabilistic methods (all except ACE, which only reports scores for its single deterministic prediction).
How PDFdigest Helps You Understand Research
Instant Paper Analysis
Get structured summaries and key findings from dense PDFs in seconds.
Visual Explanations
Turn complex methods, figures, and results into clearer visual breakdowns.
AI-Powered Q&A
Ask focused questions and get answers grounded in the paper.
Results & Findings
Climate models are foundational tools used to understand how the Earth system evolves over long time periods and how it may change as a response to possible greenhouse gas emission scenarios. As a result, scientists and policymakers are limited to exploring only a small subset of possibilities for different mitigation and adaptation strategies .
- Climate models are foundational tools used to understand how the Earth system evolves over long time periods and how it may change as a response to.
- As a result, scientists and policymakers are limited to exploring only a small subset of possibilities for different mitigation and adaptation strategies .
- Indeed, most such models only report forecasts up to two weeks into the future and may diverge or become physically inconsistent over longer simulations.
- In contrast, climate projections demand accurate and stable simulations of the global Earth system spanning decades or centuries, requiring reliable reproduction of long-term statistics.
- In Figure 1 we quantitatively show this divergence between the medium-range weather forecasting skill of ML models (measured as the average RMSE on 5-day forecasts) and.
Both these baselines are unable to outperform or even match the scores of the deterministic ACE baseline.
As a result, scientists and policymakers are limited to exploring only a small subset of possibilities for different mitigation and adaptation strategies .
Practical Applications
Training relatively cheap-to-run data-driven surrogates to emulate global climate models could provide a compelling alternative . Even then, the problem remains that due to optimizing them on MSE-based loss functions, the deterministic predictions may degrade to a mean prediction for longer forecast time scales and underestimate unlikely events .
This has important implications for practitioners, implying that optimizing for short-term forecasts alone -as is current practice for most ML-based weather forecasting models -may be suboptimal for attaining accurate climate simulations.
This is a little-discussed observation that has important implications for ML practitioners since it implies that optimizing for short-term forecasts alone -as is current practice for most ML-based weather forecasting models -may be suboptimal for attaining accurate climate simulations.
Problem Setting
The problem setting defines the goal of learning the probability distribution over a horizon of time steps, conditional on initial conditions and forcing variables. It outlines the need for training on shorter horizons and the autoregressive application of the model.
Frequently Asked Questions
Such climate simulations are currently very expensive to generate due to the computational complexity of the underlying physics-based climate models, which must be run on supercomputers. ACE was trained on a next-step forecasting objective based on a MSE loss.
\u2022 ACE-STO: We re-train ACE but use MC dropout, in the same way how it is applied in SFNO \u03d5 for our method, to generate stochastic predictions. This makes C48 around 8\u00d7 less computationally costly to run compared to the reference simulations.
ACE is trained to emulate the United States’ primary operational global forecast model, the physics-based FV3GFS , which is operationally used at the US National Weather Service and US National Centers for Environmental Prediction. While it is possible to ensemble a deterministic.
Training relatively cheap-to-run data-driven surrogates to emulate global climate models could provide a compelling alternative . This has important implications for practitioners, implying that optimizing for short-term forecasts alone -as is current practice for most ML-based weather forecasting models -may be suboptimal.
As a result, scientists and policymakers are limited to exploring only a small subset of possibilities for different mitigation and adaptation strategies . Both these baselines are unable to outperform or even match the scores of the deterministic ACE baseline.
This paper discusses a new model that helps simulate climate changes over long periods using advanced machine learning techniques. It aims to make climate predictions more accurate and less resource-intensive.