AI Surrogates Speed Up Smog Simulations by 1000x

This article provides an overview of how atmospheric chemists use artificial intelligence (AI) surrogate models to drastically accelerate chemical transport simulations. Traditional models used to predict ozone and secondary organic aerosol (SOA) smog require massive computational power to solve complex mathematical equations. By replacing these heavy physics-based calculations with trained neural networks, scientists can now run smog simulations up to a thousand times faster while maintaining high accuracy.

Traditional chemical transport models (CTMs) are the standard tools for forecasting air quality. They simulate how pollutants move, react, and transform in the atmosphere. However, accurately modeling the formation of ground-level ozone and secondary organic aerosols is notoriously slow. It requires solving stiff differential equations that represent thousands of simultaneous chemical reactions, alongside complex physics governing wind patterns, temperature, and sunlight. A single high-resolution simulation can take days or weeks on a supercomputer, limiting how quickly scientists can respond to air quality crises.

To bypass this bottleneck, atmospheric chemists use AI surrogate models, also known as emulators. Instead of calculating every individual chemical reaction from scratch, an AI surrogate learns the relationship between input conditions and the resulting pollutant concentrations.

The process begins by generating a massive dataset using traditional, physics-based CTMs. This dataset includes inputs such as:

• Precursor emission rates (nitrogen oxides, volatile organic compounds, sulfur dioxide).
• Meteorological data (temperature, humidity, wind speed, solar radiation).
• Geographic data (topography and land use).

Researchers feed these inputs, along with the corresponding ozone and SOA outputs produced by the traditional model, into deep learning architectures like Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), or Transformers. During this training phase, the AI identifies the underlying patterns and non-linear relationships that govern smog formation.

Once the AI model is trained, it acts as a direct substitute—or surrogate—for the traditional model's chemical solver. This is where the 1000x speedup occurs. While a traditional model must iteratively solve differential equations step-by-step over simulated time, a neural network relies on matrix multiplication. Modern Graphics Processing Units (GPUs) are specifically designed to execute matrix multiplications in parallel at lightning speeds. The AI takes the initial weather and emission data and predicts the final ozone and SOA concentrations in a fraction of a second, completely skipping the intermediate mathematical steps.

To ensure the AI does not produce physically impossible results, chemists often use Physics-Informed Neural Networks (PINNs). These models have physical laws, such as the conservation of mass, coded directly into their loss functions. This forces the AI to adhere to real-world chemistry constraints while maintaining its speed advantage.

This 1000-fold increase in simulation speed transforms air quality management. Scientists can now run thousands of "what-if" scenarios in minutes to see how specific emission cuts will affect urban smog. Furthermore, it allows for real-time, high-resolution air quality forecasting, giving the public and policymakers immediate warnings about dangerous ozone and particulate matter events.