🔬 physics

Acceleration of horizontal numerical advection for atmospheric modeling through surrogate modeling with temporal coarse-graining

This paper presents a machine-learned surrogate model that accelerates horizontal advection simulations in atmospheric modeling by up to 92 times through temporal coarse-graining, achieving significant speedups while maintaining acceptable accuracy and demonstrating generalization across seasons and vertical levels.

Original authors: Manho Park, Christopher V. Rackauckas, Christopher W. Tessum

Published 2026-05-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Manho Park, Christopher V. Rackauckas, Christopher W. Tessum

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: Simulating the Wind is Slow

Imagine you are trying to predict how a cloud of smoke or a scent of perfume moves through a city. Scientists use complex computer programs to do this. These programs break the city down into a giant grid of tiny squares (like a chessboard) and calculate how the wind pushes the smoke from one square to the next every few seconds.

The problem is that this is incredibly slow. To get an accurate picture, the computer has to take tiny, cautious steps. It's like trying to walk across a room by taking steps the size of a grain of sand. You will get there eventually, but it will take forever. This slowness stops scientists from running many simulations at once or making very detailed predictions quickly.

The Old Way vs. The New Way

The Old Way (Spatial Coarse-Graining):
Previously, researchers tried to speed this up by making the "chessboard" bigger. Instead of tiny squares, they used huge squares.

The Analogy: Imagine looking at a high-resolution photo of a forest. To make it load faster, you blur it until the trees look like green blobs. It loads instantly, but you can no longer see individual leaves or birds.
The Issue: In air quality modeling, losing that detail is bad. If you blur the grid, you might miss a dangerous pocket of pollution in a specific neighborhood.

The New Way (Temporal Coarse-Graining):
This paper introduces a new method that keeps the "chessboard" tiny and detailed (so you don't lose resolution) but changes how often the computer checks the position of the smoke.

The Analogy: Imagine you are watching a movie of a car driving.
- Standard method: You watch every single frame (60 times a second). It's smooth but takes a long time to process.
- This new method: You use a "smart AI" that watches the first few frames, figures out the pattern, and then jumps ahead to see where the car will be 10 seconds later. It skips the boring middle frames but still knows exactly where the car is.
- The Catch: The AI needs to be trained to be good at guessing those jumps.

How They Built the "Smart AI"

The researchers built a machine-learning solver (a type of AI) that acts as a super-fast traffic cop for the wind.

The Training: They taught the AI using wind data from January in North America. They showed it how the wind moves pollutants over short periods.
The Trick: They trained the AI to take "big steps" in time. Instead of calculating the wind every 5 minutes, they trained it to calculate where the wind would be 20, 40, or even 160 minutes later in a single step.
The Result: The AI learned to predict the movement of air without needing to check every single tiny second.

What They Found

The researchers tested this AI to see if it could handle things it hadn't seen before, like different seasons or wind high up in the sky.

Speed vs. Accuracy Trade-off: They found a clear rule: The faster you want the simulation to go, the slightly less accurate it becomes.
- They could make the simulation 92 times faster while still keeping it about 60% accurate.
- If they wanted it 10 times faster, it was still about 98% accurate.
- Think of it like a car: You can drive 100 mph (very fast) but you might miss a small pothole. If you drive 10 mph, you see every pebble. This AI lets you choose your speed.
Generalization (The "Unseen" Test):
- Seasons: The AI was trained only on January data. When they tested it on other months, it worked great for most of the year. However, it got "confused" and unstable during June and October. It seems the wind patterns in those months were too different from January for the AI to guess correctly without help.
- Altitude: The AI was trained on ground-level wind. It worked surprisingly well for high-altitude winds (up to 21 km high) when taking small "time jumps." But when they tried to make the time jumps very large (32x faster), the AI failed at high altitudes.

Why This Matters

This study proves you don't have to sacrifice the "zoom level" (spatial resolution) to get speed. You can keep the detailed map of the city and just make the computer "think" faster about the future.

The authors suggest this could be useful for:

Screening tools: Quickly checking if a pollution event is a big deal before running a slow, detailed simulation.
Ensemble simulations: Running the same model 100 times with slightly different starting conditions to get a better forecast (which is usually too slow to do with standard methods).
Data Assimilation: Updating weather models in real-time with new data.

The Bottom Line

The researchers built a "time-jumping" AI that simulates wind and pollution movement up to 92 times faster than standard methods without blurring the map. It works best when the weather is similar to what it was trained on, and it offers a flexible trade-off: you can choose to go faster if you are willing to accept a tiny bit less precision.

Technical Summary: Acceleration of Horizontal Numerical Advection via Temporal Coarse-Graining

Problem Statement
Numerical advection is a computationally expensive operation in geoscientific models, often serving as the most or second-most costly module in systems like E3SM, CMAQ, GEOS-Chem, and CAMx. High computational costs limit the scale and resolution of simulations. While reducing spatial resolution (spatial coarse-graining) can accelerate simulations, it degrades the relevance of outputs for applications like air quality modeling, where fine spatial detail is critical to accurately assess pollution exposure and disparities. Furthermore, existing machine learning (ML) approaches that achieve speedups often rely on spatial coarse-graining or hardware acceleration, or they fail to produce stable simulations when applied to original spatial resolutions without spatial reduction. Additionally, traditional numerical solvers are constrained by the Courant-Friedrichs-Lewy (CFL) condition, which limits the time step size based on flow velocity and grid spacing to ensure stability.

Methodology
The authors propose a machine-learned advection solver that achieves computational acceleration through temporal coarse-graining alone, preserving the original spatial resolution. The core innovation is training the model to take integration steps larger than those dictated by the CFL condition.

Solver Architecture: The solver is a Convolutional Neural Network (CNN) designed to emulate a finite-volume flux-form approach. It takes the concentration and CFL number of a grid cell and its neighbors as input and outputs the mass flux across the cell's left boundary. The architecture consists of three hidden layers with 10 parallel convolutional filters per layer. The first two layers use GeLU activation, and the final layer uses a hyperbolic tangent activation to ensure output symmetry.
Information Horizon: Unlike traditional solvers restricted by local gradients, the ML solver's "information horizon" (the number of grid cells influencing a single step) expands with the temporal coarsening factor. The authors increase the convolutional kernel size as the coarsening factor increases to capture this expanded horizon.
Training Data: The model was trained on 1-D advection data generated using ground-level wind fields over North America (January 2018) using the Piecewise Parabolic Method (PPM) from the GEOS-Chem Classic model. The training dataset included 10-day simulations with a 5-minute time step.
Temporal Coarsening Strategy: To train for larger time steps, the authors downsampled the training data to create datasets with coarsening factors of 4×, 8×, 12×, 16×, and 32×. Crucially, they accumulated flux values over the coarsened time steps to ensure mass conservation, training the model to predict the sum of fluxes over multiple steps rather than a single instantaneous flux.
Implementation: The model was trained using Flux.jl and transferred to SimpleChains.jl, a framework optimized for small neural networks using stack-allocated operations, to achieve significant inference speedups (5–10× faster than standard deep learning packages).
Evaluation: Performance was evaluated on 2-D advection simulations (using non-directional splitting) over the same domain. Generalization was tested against unseen data: wind fields from different months (seasonal generalization) and wind fields from 72 different vertical levels (vertical generalization).

Key Results

Speed-Accuracy Tradeoff: The solvers demonstrated a favorable tradeoff between speed and accuracy. The model achieved a maximum 92× speedup compared to the reference solver while maintaining an $r^2$ of 0.60. A linear regression analysis revealed that for every factor of 10 increase in speed (within the 8×–32× coarsening range), the model incurred an $r^2$ loss of approximately 0.26.
Generalization Performance:
- Vertical Levels: The 4×-coarsened solver successfully reproduced simulations across all 72 vertical levels. Solvers with 8×–16× coarsening performed well up to the 40th level (~~21 km), while the 32× solver was limited to the first 10 levels (~~1.3 km).
- Seasonality: The solvers generalized well across most months of 2018 when trained on January data. However, instabilities were observed in June and October, particularly for higher coarsening factors.
Stability and CFL Limits: The instability in June and October, as well as at higher vertical levels, correlated with maximum CFL values exceeding those present in the training dataset (maximum training CFL was 0.386). The solver struggled to maintain numerical stability when the effective CFL (training CFL × coarsening factor) significantly exceeded the training range.
Mass Conservation: While designed to be mass conservative, the solver exhibited slight mass increases at the start of simulations due to the clipping of negative concentration predictions to zero (a positivity limiter). This resulted in a net mass increase before mass loss occurred via lateral boundaries.

Significance and Claims
The paper claims to demonstrate a novel approach to accelerating geoscientific advection simulations without sacrificing spatial resolution, addressing a major limitation of previous ML-based fluid dynamics studies. By leveraging temporal coarse-graining, the authors show it is possible to bypass the CFL constraint for inference, offering a tunable tradeoff between computational speed and accuracy.

The authors position this work as a foundational step toward operational use. They acknowledge that the current models are not yet optimized for production use, noting limitations such as instability under high CFL conditions (outside the training distribution) and the violation of monotonicity due to negative concentration clipping. They suggest that with additional fine-tuning, diverse training data (including different seasons and vertical levels), and hyperparameter optimization, these solvers could be suitable for screening tools, ensemble simulations, or data assimilation where trading some accuracy for significant speed is advantageous. The study emphasizes that this approach allows for the potential to increase spatial resolution (by saving computational time) rather than decreasing it, aligning with the goals of modern atmospheric modeling.