Machine Learning of Vertical Fluxes by Unresolved Midlatitude Mesoscale Processes

This study demonstrates that machine learning can effectively parameterize unresolved midlatitude mesoscale vertical fluxes in Earth system models by leveraging vertically non-local information from temperature, moisture, and meridional wind, thereby offering a targeted approach to improve the representation of complex processes like slantwise convection and frontal dynamics.

The Gist

Imagine you are trying to predict the weather, but you are looking at the atmosphere through a very blurry, low-resolution camera. In this blurry picture, you can see big storm systems and general wind patterns, but you miss the tiny, chaotic details: the swirling eddies, the narrow bands of rain, and the sudden bursts of heat that happen between the big clouds.

In climate science, these missing details are called mesoscale processes. They are too small for standard climate models to "see," but they are powerful enough to change the big picture.

This paper is about teaching a computer (using Machine Learning) to guess what those missing details look like, so the blurry camera can be "fixed" without needing a super-expensive, high-resolution camera.

Here is the story of how they did it, explained simply:

1. The Problem: The "Blurry" Climate Model

Think of a standard climate model like a low-resolution video game. The graphics are blocky. You can see the mountains and the ocean, but you can't see the individual trees or the way wind swirls around a specific tree.

• The Issue: In the mid-latitudes (like the North Atlantic near the Gulf Stream), there are sneaky weather processes called slantwise convection and frontal dynamics. These are like invisible elevators moving heat and moisture up and down the atmosphere. Because the model is "blurry," it misses these elevators.
• The Consequence: Without these elevators, the model gets the temperature and wind wrong, especially over the ocean where cold air hits warm water.

2. The Solution: The "High-Def" Training Video

To teach the computer how to fix the blur, the researchers needed a "High-Definition" reference video.

• They ran a super-detailed simulation of the North Atlantic (using a grid so fine it could see the "trees").
• From this high-def video, they calculated exactly how much heat and moisture were moving up and down in those tiny, invisible elevators.
• The Goal: They wanted to train an Artificial Neural Network (a type of AI) to look at the "blurry" version of the weather and predict what the "high-def" elevators were doing.

3. The Experiment: Teaching the AI

They treated the AI like a student taking a test.

• The Input (The Clues): They gave the AI the "blurry" weather data: temperature, wind, and humidity at different heights.
• The Output (The Answer): They asked the AI to guess the vertical movement of heat and moisture.
• The Twist: They didn't just want the AI to guess; they wanted to know how it guessed. Did it just memorize the answer, or did it understand the physics?

4. The Big Discoveries (The "Aha!" Moments)

A. You Can't Just Look at One Spot

The researchers found that to guess what's happening in one column of air, the AI needs to look at its neighbors.

• The Analogy: Imagine trying to guess if it's raining in your backyard. If you only look at your own grass, you might miss it. But if you look at the wet grass in your neighbor's yard and the dark clouds moving from the north, you can guess it's raining.
• The Finding: The AI learned that cold air outbreaks (cold wind blowing from the north over warm ocean water) create these invisible elevators. The AI needed to see the "big picture" of the cold air mass to predict the local turbulence.

B. The "Vertical Velocity" Trap

One of the most important clues the AI used was vertical velocity (how fast air is moving up or down).

• The Trap: The AI got really good at predicting the weather when it was allowed to use this clue. However, the researchers realized this clue is a "cheat."
• Why? In the real world (or in a standard low-resolution model), you can't see how fast the air is moving up or down directly. You have to guess it. The AI was using a "superpower" it wouldn't have in a real climate model.
• The Lesson: If you include this cheat in a real model, the model might crash or become unstable. The researchers learned that while vertical velocity is a great predictor, it's a dangerous one to use in the final product.

C. The "Cold Air Outbreak" Connection

The AI discovered a strong link between Cold Air Outbreaks (when frigid air sweeps over the warm Gulf Stream) and the invisible elevators.

• The Mechanism: When cold, dry air hits warm, moist ocean water, it gets unstable. It wants to rise. This creates those slantwise convection bands.
• The AI's Insight: The AI learned that if it sees cold air near the surface and warm air above, it should predict a massive upward rush of heat and moisture. This helps explain why the Gulf Stream region is so stormy.

5. The Takeaway: Why This Matters

This paper is a roadmap for the future of climate modeling.

• We need more clues: You can't just use a few simple numbers to predict these complex weather patterns. You need a lot of data (temperature profiles, wind shear, humidity) to get it right.
• Non-local thinking: Weather isn't just about what's happening right here; it's about what's happening nearby and above. The AI learned to think in 3D and look at neighbors.
• Better Models: By using these AI tricks, future climate models can be more accurate without needing to run on supercomputers that cost millions of dollars. They can "hallucinate" the missing details correctly.

In a nutshell: The researchers taught a computer to fill in the missing details of a blurry weather map by learning from a high-definition movie. They discovered that the computer needed to look at its neighbors and understand the drama of cold air hitting warm water to get the job done, but they also learned to be careful not to give the computer "cheat codes" that wouldn't work in the real world.

Technical Summary

Here is a detailed technical summary of the paper "Machine Learning of Vertical Fluxes by Unresolved Midlatitude Mesoscale Processes."

1. Problem Statement

Earth System Models (ESMs) typically operate at coarse resolutions (e.g., ~100 km) that cannot explicitly resolve midlatitude mesoscale processes (scales < 500 km), such as slantwise convection, frontal dynamics, and conditional symmetric instability (CSI). These processes are critical for vertical transport of momentum, heat, and moisture, particularly in regions like the Gulf Stream. While Machine Learning (ML) has been successfully applied to parameterize deep convection in idealized settings (e.g., aquaplanets), its application to realistic midlatitude mesoscale fluxes remains underexplored.

The specific challenges addressed in this study include:

• Determining if ML can learn vertical flux profiles resulting from complex, non-local midlatitude dynamics.
• Identifying which atmospheric state variables are most informative for predicting these fluxes.
• Understanding the vertical localization relationships (local vs. non-local dependencies) required for accurate parameterization.
• Assessing the "representativeness" of coarse-grained variables (like vertical velocity) derived from high-resolution data versus those generated by low-resolution models.

2. Methodology

Data Source and Setup

• Simulation: The study utilizes a variable-resolution simulation from the Community Earth System Model version 2.2 (CESM2.2) with the Community Atmospheric Model (CAM6).
• Resolution: The grid is refined to 14 km over the North Atlantic (specifically the Gulf Stream region), while the global background remains at ~110 km. This allows the model to resolve mesoscale features (like slantwise convection) while keeping deep convection parameterized.
• Dataset: 30 years of 6-hourly data were subsampled to capture diurnal and seasonal cycles, resulting in ~10 million training samples.
• Target Region: The Gulf Stream region (25°N–55°N, 90°W–30°W).

Feature Engineering and Targets

• Targets: The model predicts vertical profiles of subgrid-scale vertical fluxes for four variables: Zonal wind ($U$), Meridional wind ($V$), Temperature ($T$), and Moisture ($Q$). These are calculated as the covariance between the variable and vertical velocity ($\omega$) after conservative remapping to a coarse grid (~100 km).
• Inputs (Features):
  • Vertical profiles of $T, U, V, \omega, Q$.
  • Scalar variables: Surface pressure ($P_s$) and Convective Available Potential Energy (CAPE).
  • Non-local information: Horizontal derivatives ($\partial_x, \partial_y$) of $U, V, T$ to capture neighboring column information (critical for tilted motions like slantwise convection).
  • Total: 244 input features mapping to 88 output targets (vertical profiles).

Machine Learning Architecture

• Model: A feedforward Artificial Neural Network (ANN) with 8 hidden layers, batch normalization, and GELU activation functions.
• Loss Function: Balanced L1 loss was selected to handle the heavy-tailed distribution of flux data, reducing sensitivity to outliers compared to standard MSE.
• Training: Optimized using Optuna for hyperparameter tuning, with early stopping based on validation loss.
• Explainability: The study employs Ablation Experiments (removing feature categories) and SHAP (SHapley Additive exPlanations) values to interpret feature importance and physical mechanisms.

Regime Definition

To analyze generalizability, the authors used K-means++ clustering to define five distinct atmospheric regimes based on wind shear, temperature gradients, moisture, and flux characteristics:

1. Stable: High pressure, low fluxes.
2. Cold Air Outbreak (CAO): Dry, cold air over warm surfaces.
3. Precipitating Convection: High CAPE, heavy resolved precipitation.
4. Organized Heavy Precipitation: Extreme fluxes, high CAPE, strong convergence.
5. Warm Moist Air Masses: Warm, moist conditions.

3. Key Results

Model Performance

• Skill: The ANN achieved $R^2$ scores between 0.5 and 0.8 for most fluxes in the troposphere, significantly outperforming Multiple Linear Regression (MLR) baselines, particularly for momentum fluxes.
• Regime Dependence: Performance varied by regime. The model struggled most with the rare "Organized Heavy Precipitation" regime (Cluster 3), where extreme values were harder to predict, though sign prediction remained relatively robust.
• Vertical Localization:
  • Profile-to-Profile mapping (using full vertical profiles) performed best.
  • Neighbor-to-Level (using adjacent levels) performed significantly better than Level-to-Level (local only), confirming the importance of non-local vertical dependencies.

Feature Importance (Ablation & SHAP)

• Critical Features:
  • Vertical Velocity ($\omega$): When included, it provided the largest single boost to skill (up to ~20% improvement in profile-to-profile, and massive gains in local mappings).
  • Horizontal Divergence: Crucial for predicting all flux types.
  • Temperature Profiles: Essential for capturing stability and non-local effects.
  • Moisture: Highly important for moisture and heat fluxes.
• The "Vertical Velocity" Paradox: The study found that coarse-grained vertical velocity from high-resolution simulations contains information about subgrid-scale events that is not present in vertical velocity generated by a true low-resolution model. Consequently, while $\omega$ is a powerful predictor in offline training, its inclusion in an online ESM parameterization could lead to instabilities or errors because the coarse-resolution model cannot reproduce the specific $\omega$ values the ML model expects.

Physical Insights (SHAP Analysis)

• Non-Local Effects: The model learned that surface conditions (temperature, moisture, wind) strongly influence fluxes throughout the entire boundary layer, not just at the surface.
• Cold Air Outbreaks (CAOs): The analysis confirmed that cold, dry northerly winds over warm oceans trigger strong upward heat and moisture fluxes, consistent with slantwise convection mechanisms.
• Stability vs. Instability: The model identified that enhanced stability (often associated with cold fronts) can suppress upright convection but facilitate slantwise convection (CSI), a nuance captured by the non-linear interactions in the ANN.

4. Significance and Contributions

1. Targeted Parameterization: This work demonstrates that ML can successfully parameterize complex midlatitude mesoscale processes (slantwise convection, fronts) that are distinct from deep convection, moving beyond idealized aquaplanet setups.
2. Importance of Non-Local Information: The study highlights that accurate parameterization of midlatitude fluxes requires vertical profile information and neighboring column data (horizontal derivatives). Simple local mappings are insufficient.
3. The Vertical Velocity Caveat: A major contribution is the identification that coarse-grained vertical velocity acts as a "proxy" for unresolved dynamics in high-res data but is not a valid predictor for low-resolution models. The authors recommend excluding $\omega$ from online parameterizations to ensure stability, relying instead on thermodynamic and dynamic state variables.
4. Explainability in Climate ML: By combining ablation studies with SHAP values, the authors provide a physical interpretation of the "black box" model, linking ML predictions to known physical mechanisms like CSI and frontal dynamics.
5. Guidance for Future ESMs: The results suggest that improving ESMs with ML requires:
   • Including a large number of features (profiles + derivatives).
   • Focusing on dry variables (which explain a large portion of the flux variance) alongside moisture.
   • Avoiding direct reliance on coarse-grained vertical velocity for online coupling.

Conclusion

The paper establishes that ML parameterizations can effectively capture midlatitude mesoscale vertical fluxes, provided they utilize non-local vertical profiles and horizontal derivatives. However, the successful offline performance relies partly on features (like coarse-grained $\omega$) that may not be physically consistent in a coupled, low-resolution climate model. The study provides a roadmap for developing stable, physics-informed ML parameterizations that account for the complex, multi-scale interactions of the midlatitude atmosphere.