Emulating the Forced Response of Climate Models with Flow Matching

This paper presents a deep learning flow matching model that efficiently emulates the forced response of global climate models to various Shared Socioeconomic Pathways and simultaneous climate forcings, successfully generating unseen scenarios and demonstrating the necessity of including diverse forcings to accurately capture long-term atmospheric trends.

The Gist

Imagine trying to predict the weather for the next 85 years. You have a supercomputer running a massive, incredibly detailed simulation of the Earth's atmosphere, oceans, and land. This is what scientists call a Climate Model. It's like a giant, digital twin of our planet.

The problem? Running this digital twin is incredibly slow and expensive. It takes a supercomputer days or weeks to simulate just a few decades. If policymakers want to know what happens if we cut emissions by 50% versus 100%, they need to run hundreds of these simulations to get a clear picture. But we can't wait that long.

The Solution: A "Climate Co-Pilot"
This paper introduces a new tool: a Deep Learning Emulator. Think of this not as a replacement for the supercomputer, but as a highly trained "co-pilot" or a "speed-demon" version of the climate model.

The researchers taught an AI to watch the supercomputer run simulations and learn its "personality." Once trained, this AI can generate future climate scenarios in seconds that look and feel almost exactly like the slow, expensive supercomputer runs.

How It Works: The Recipe Analogy

To understand how this AI learns, imagine you are trying to teach a robot chef to bake a cake that changes flavor based on the ingredients you give it.

1. The Ingredients (Forcings): In the climate world, the "ingredients" are things like Carbon Dioxide (CO2), Methane, Ozone, and tiny dust particles called Aerosols. These are the external drivers that change the climate.
2. The Recipe (The Model): The AI is the chef. It needs to know how the cake (the Earth's climate) reacts when you add more sugar (CO2) or a pinch of salt (Aerosols).
3. The Training: The researchers fed the AI thousands of "batches" of cake made by the real supercomputer, showing it exactly what happened when different amounts of ingredients were added.

The Big Discovery: Not All Ingredients Are Created Equal

The most interesting part of this paper is what happened when the researchers tried to bake the cake with missing ingredients. They ran experiments where they told the AI to ignore certain ingredients to see if it still worked.

• The "Sugar" Test (Greenhouse Gases): When they removed the greenhouse gases (like CO2) from the AI's instructions, the chef completely failed. The cake didn't get hotter over time. The AI couldn't predict the long-term warming trend. Lesson: For this specific monthly-resolution setup, you absolutely need the greenhouse gas data to predict the future climate.
• The "Dust" Test (Aerosols): Aerosols are tiny particles (like pollution or volcanic ash) that actually cool the Earth by reflecting sunlight. The researchers found something surprising: when they removed the aerosol data, the AI actually baked a better cake. It was more accurate and stable.
  • Why? The paper suggests that aerosols are like "noisy" ingredients. They change very quickly and randomly (like a chaotic sprinkler). Because the AI only looks at monthly averages, the aerosol data looked like static noise rather than a clear signal. It confused the chef.
• The "Sky Structure" Test (Ozone): Ozone is a gas high in the sky that acts like a structural beam for the atmosphere. When they removed ozone, the AI's simulation collapsed. It couldn't figure out how the temperature changed from the ground up to the stratosphere. Lesson: In this particular architecture, ozone is essential for the AI to understand the vertical structure of the sky.

The "Overshoot" Challenge

The researchers also tested the AI on a tricky scenario called an "Overshoot." Imagine a world where we heat up the planet, then suddenly try to cool it down by sucking CO2 out of the air.

• The AI was trained on scenarios where things just got hotter and hotter.
• They asked the AI to predict this "cooling down" scenario, which it had never seen before.
• Result: Even though the AI was trained only on warming scenarios, it performed essentially as well on the overshoot scenario as it did on the warming-only ones. This is a good sign — it suggests the AI has learned the underlying physics of the climate response, not just memorized a simple "warming trend" pattern.

The Comparison: AI vs. The Old Way

The team compared their new AI to an existing tool called MESMER-M.

• MESMER-M is incredibly fast at generating ensembles—actually faster than this new AI. It is not rigid; it is a highly effective tool for its specific purpose.
• However, MESMER-M has limitations: It only predicts land-surface temperature, whereas this new AI covers both land and ocean with full vertical atmospheric structure. Additionally, MESMER-M requires global yearly-mean values as inputs and is not autoregressive (it cannot roll its own state forward step-by-step).
• The New AI's Advantage: The strength of this new tool isn't raw ensemble speed, but its broader coverage (land + ocean, full vertical structure, monthly resolution) and its autoregressive flexibility, allowing it to simulate complex, evolving states over time.

The Bottom Line

This paper shows that we can build a fast, AI-powered "climate co-pilot" that mimics the slow, expensive supercomputers. However, to make it work, we have to be very careful about what data we feed it:

1. Must-haves (for this setup): In this specific monthly-resolution emulator with this architecture, greenhouse gases and ozone are non-negotiable; without them, the AI fails to predict the future.
2. Maybe-haves: Aerosols (pollution particles) might actually be too messy for this specific type of AI to handle well right now, and leaving them out might make the predictions more accurate.

The goal isn't to replace the supercomputers, but to give scientists a tool that can run thousands of simulations instantly, helping them make better decisions about our planet's future.

Technical Summary

Technical Summary: Emulating the Forced Response of Climate Models with Flow Matching

Problem Statement

Global climate models, specifically Earth System Models (ESMs), are essential for simulating past and future climate pathways under Shared Socioeconomic Pathways (SSPs). However, running these models is computationally prohibitive, limiting the generation of large ensembles required to robustly estimate internal variability and scenario uncertainty. While recent machine learning (ML) approaches have shown promise in emulating climate dynamics, a significant gap remains in accurately modeling non-stationary, long-term forced responses to external drivers (forcings) such as greenhouse gases, aerosols, and ozone. Existing ML emulators often rely on pre-determined ocean conditions or fail to capture the shifting distributions of climate states driven by complex forcing combinations. This work addresses the challenge of training a Deep Learning (DL) emulator capable of generating unseen future climate scenarios that respond dynamically to a multitude of simultaneous external forcings.

Methodology

The authors propose ArchesClimate-SSP, an extension of the ArchesClimate framework (itself based on ArchesWeather/Pangu-Weather), utilizing Flow Matching as a generative mechanism.

Data and Forcings

• Source Data: The model is trained on outputs from the IPSL-CM6A-LR ESM, covering 8 future SSP scenarios (ranging from 1.9 to 8.5 W/m² radiative forcing), historical runs, and control runs (2015–2100).
• Forcings: The model ingests a diverse set of exogenous forcings:
  • Greenhouse Gases (GHGs): CO₂, CH₄, and N₂O (sourced from Input4MIPs).
  • Aerosols: Four types (ASNO3M, CSNO3M, CINO3M, SO₄) derived from the INCA chemistry model.
  • Ozone: Mole fractions across 10 vertical layers (troposphere to mesosphere).
  • Solar Spectral Irradiance (SSI): Non-spatial monthly data.
• Preprocessing: Spatial forcings (ozone, aerosols, GHGs) are regridded to the IPSL resolution (2.5°×1.25°) and concatenated to the input state. Non-spatial forcings (global GHG means, SSI) are integrated via Conditional Layer Normalization.

Architecture and Training

• Model Structure: A Sliding Window Vision Transformer processes spatial tokens. The prediction task is split into a deterministic mean prediction and a generative residual to capture probabilistic uncertainty.
• Training Modifications:
  • Forcing Dropout: To prevent overfitting to specific forcing signals and encourage the model to internalize climate dynamics, forcing inputs are stochastically masked with probability $\lambda = 0.8$.
  • Randomized Time Horizons: The model is trained on random future timesteps ($h \in \{1, 6, 12\}$ months) rather than just the immediate next step to enhance long-term stability.
  • Loss Function: A composite loss combines Mean Squared Error (MSE) for the deterministic mean and a spectral loss for the generative component to improve variability.

Experimental Design

The study employs an ablation strategy to test the necessity of specific forcings. Models are trained with:

1. ACfull: All forcings included.
2. AC¬ghg: Greenhouse gases removed.
3. AC¬aero: Aerosols removed.
4. AC¬o3: Ozone removed (failed to converge/stabilize).
   The models are evaluated against the IPSL ground truth and the statistical emulator MESMER-M on unseen scenarios (SSP4-3.4, SSP5-3.4, SSP5-8.5) and an abrupt-4xCO₂ experiment.

Key Contributions

1. Novel Training Mechanisms: Introduction of forcing dropout and randomized time horizon sampling to improve long-term stability and flexibility in generative climate emulators.
2. Unseen Scenario Generation: Demonstration that a DL model can accurately reproduce the temporal and spatial dynamics of SSP scenarios not seen during training.
3. Forcing Sensitivity Analysis: Empirical evidence showing that the explicit inclusion of a specific range of forcings is critical for accuracy. Specifically, the study identifies that while GHGs and ozone are essential for trend stability and vertical structure, the inclusion of complex aerosol data can degrade performance due to high-frequency noise.

Results

• Scenario Emulation: ArchesClimate-SSP successfully emulates unseen SSP scenarios (e.g., SSP4-3.4, SSP5-3.4). The AC¬aero model (without aerosols) achieved the lowest Root Mean Square Error (RMSE) for land surface temperature, outperforming both the full model and the GHG-ablated model, and performing comparably to the state-of-the-art statistical baseline MESMER-M.
• Forcing Necessity:
  • GHGs: Removing GHG forcings (AC¬ghg) led to a progressive increase in error over time, particularly in the tropics, confirming that GHGs are the primary anchor for capturing secular warming trends.
  • Ozone: The ozone-ablated model failed to produce stable checkpoints, highlighting ozone's critical role in resolving the vertical thermal structure of the atmosphere.
  • Aerosols: Contrary to the assumption that more data is better, removing aerosols (AC¬aero) improved accuracy. The authors hypothesize that the high spatial and temporal variability of aerosols (e.g., aerosol-cloud interactions) introduces stochastic noise that confounds the monthly-resolution learning process, whereas GHGs and ozone provide more stable, physically coherent signals.
• Physical Consistency: The models maintained hydrostatic balance (relationship between temperature and geopotential height) across the vertical column, with residuals near zero, indicating physical consistency in the learned representations.
• Extrapolation: In the extreme SSP5-8.5 scenario (out-of-distribution), the models maintained numerical stability but exhibited biases, particularly in the density distribution of global temperatures, underscoring current limitations in extreme extrapolation.

Significance and Claims

The paper claims that ArchesClimate-SSP offers a robust method for quickly generating spatially and temporally explicit future climate simulations that respond appropriately to external forcings. Its significance lies in:

• Efficiency: Providing a computationally cheaper alternative to running full ESM ensembles for multiple scenarios.
• Ensemble Generation: The ability to generate diverse ensemble members from single ESM runs, enabling probabilistic analysis for scenarios that originally lacked multiple realizations.
• Physical Grounding: Demonstrating that a generative framework can capture shifting atmospheric dynamics across the troposphere and stratosphere when provided with the correct set of forcings.
• Feature Selection Insight: The work challenges the assumption that including all available forcing data maximizes accuracy, suggesting that for monthly-resolution emulators, simplified or filtered representations of forcings (specifically regarding aerosols) may yield better predictive skill by reducing variance and noise.

The authors conclude that while the model successfully captures long-term trends and vertical structure, challenges remain in resolving high-variance variables (like wind) and inter-annual variability, which are necessary for capturing extreme events.