Accelerating Local Air Pollution Modeling with Earth-2 CorrDiff

Traditional physics-based models struggle with the math required to track multi-component chemical reactions in the atmosphere. These CPU-heavy simulations take hours or days to run, forcing researchers to make a frustrating trade-off between spatial resolution and processing speed.

By swapping out traditional physical calculations for generative deep learning architectures, developers can now deploy high-resolution air quality mapping pipelines that run in seconds rather than hours.

Architectural Breakdown: The Two-Step Neural Network Pipeline

At the core of this approach is Earth-2 CorrDiff, an 80-million parameter neural network featuring six encoder and six decoder layers. Instead of trying to solve the entire spatial mapping problem in one complex step, the system splits the task into a two-part pipeline.

[Low-Resolution Input Data] 
             |
             v
+-----------------------------+
|   Stage 1: UNet Regression  |  <-- Resolves baseline macro-patterns
+-----------------------------+
             |
             v
   [Mean Average Prediction]
             |
             v
+-----------------------------+
|    Stage 2: AI Diffusion    |  <-- Generates localized micro-details
+-----------------------------+
             |
             v
[High-Resolution Output Map]

• The UNet Regression Stage: The pipeline first ingests low-resolution baseline data. The UNet-based regression model maps out the foundational spatial baseline across the target area, resolving stable, macro-level patterns of pollution distribution.
• The Diffusion Stage: This step acts as a generative correction layer. It evaluates the variances between the baseline macro-prediction and micro-level realities, utilizing a diffusion process to generate high-frequency localized details, sharp boundaries, and highly specific variations.

Dividing the spatial processing into two distinct steps prevents multi-scale gradient confusion during training. This pipeline finishes execution thousands of times faster than CPU-based physics solvers while reducing total energy consumption.

Model Training and Edge Hardware Deployment

Adapting foundational meteorological models to track chemical pollutants required a targeted retraining workflow:

• Training Dataset: The model was retrained on one year of historical air pollution telemetry from the United Kingdom, recorded at hourly intervals.
• Supercomputing Cluster Training: Training operations were executed on the Isambard-AI supercomputer in Bristol, which houses 5,448 NVIDIA GH200 Grace Hopper Superchips. The training run completed in 48 hours using a single eight-GPU node.
• Desktop Inference Deployment: Once trained, the optimized model weights are compact enough to run local inference. Developers deployed the model on a desktop-class DGX Spark system powered by a GB10 Grace Blackwell superchip.
• Sensor Logic Integration: For live deployments, the pipeline links with Earth-2 StormCast to ingest real-time data streams from regional ground sensors and satellite telemetry to output hourly forecasting updates.

Operational Use Cases and Open-Source Release

The localized inference speed allows researchers to configure custom scenarios to simulate how changing specific policy variables, such as modifying traffic patterns or factory emission caps, affects downwind pollution plumes.

Healthcare networks can integrate these high-resolution models to identify localized pollution spikes up to 24 hours in advance, triggering automated alerts to respiratory patients. The developers plan to release this software architecture to the open-source community, enabling regions to train local models using their own sensor networks.