🤖 machine learning

Collocation-based Robust Physics Informed Neural Networks for time-dependent simulations of pollution propagation under thermal inversion conditions on Spitsbergen

This paper proposes a robust, collocation-based Physics-Informed Neural Network (PINN) framework for simulating time-dependent pollution propagation from moving sources, demonstrating through a case study in Spitsbergen that thermal inversion significantly worsens air quality by trapping pollutants near the ground.

Original authors: Leszek Siwik, Maciej Sikora, Natalia Leszczyńska, Tomasz Maciej Ciesielski, Eirik Valseth, Manuela Bastidas Olivares, Marcin Łoś, Tomasz Służalec, Jacek Leszczyński, Maciej Paszyński

Published 2026-04-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Leszek Siwik, Maciej Sikora, Natalia Leszczyńska, Tomasz Maciej Ciesielski, Eirik Valseth, Manuela Bastidas Olivares, Marcin Łoś, Tomasz Służalec, Jacek Leszczyński, Maciej Paszyński

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 "Invisible Blanket" Problem: How Scientists are Using AI to Track Pollution in the Arctic

Imagine you are standing in a deep, narrow valley in the Arctic. It’s freezing, the sun barely rises, and you’re surrounded by mountains. Suddenly, a snowmobile zooms past you. You see a puff of exhaust, but then it seems to just... hang there. It doesn't blow away; it stays right at your feet, getting thicker and thicker.

This paper explains why that happens and introduces a high-tech "digital detective" to help predict it.

1. The Problem: The "Thermal Inversion" Trap

Normally, air behaves like a hot air balloon: warm air rises, carrying pollution up into the sky where it can be blown away by the wind.

However, in places like Spitsbergen (an island in the Arctic), something strange happens called a Thermal Inversion. Think of it like an invisible, heavy blanket made of cold air that settles into the valley. This blanket sits on top of the ground and prevents the warm, polluted air from rising. Instead of escaping, the pollution from snowmobiles and power plants gets trapped right where people breathe.

2. The Tool: The "Digital Detective" (CRVPINN)

To understand how this "blanket" moves and how much pollution it traps, scientists need to run complex simulations. Traditionally, this is like trying to predict the exact path of every single drop of ink dropped into a moving river—it’s incredibly difficult and takes massive amounts of computer power.

The researchers created a new kind of Artificial Intelligence called CRVPINN.

Here is the analogy:

Standard AI (PINN) is like a student trying to pass a test by just memorizing the answers. They might get the answers right, but they don't actually understand the logic of the math. If you ask them a slightly different question, they fail.
The New AI (CRVPINN) is like a student who is forced to learn the laws of physics to pass. It doesn't just look at the data; it understands the "rules of the game" (like how wind moves and how chemicals spread).

Because this AI understands the "rules," it is much more robust. In science-speak, "robust" means that when the AI says, "The pollution level is X," you can actually trust that it’s accurate, rather than just a lucky guess.

3. The Proof: Real-World Smoke vs. Computer Models

The scientists didn't just stay in the lab; they went into the field. They attached high-tech sensors to snowmobiles and drove them through the valleys of Longyearbyen.

The results were a "double check":

The Sensors: The real-world sensors showed massive spikes in pollution (like PM2.5, tiny particles that get deep into your lungs) whenever a snowmobile accelerated.
The AI: The CRVPINN simulation showed the exact same pattern—pollution getting trapped near the ground due to the "cold air blanket."

When the real-world measurements matched the computer's "digital twin," the scientists knew their new AI method worked.

Why does this matter?

By using this "Digital Detective," we can predict "bad air days" before they happen. If we know the "invisible blanket" is about to settle in the valley, local authorities can take action to protect people's health—perhaps by limiting snowmobile traffic or warning sensitive groups to stay indoors.

In short: They’ve built a smarter, faster, and more reliable way to see the invisible dangers in the air we breathe.

Technical Summary: Collocation-based Robust Physics Informed Neural Networks for Pollution Propagation

1. Problem Statement

The paper addresses the challenge of simulating the time-dependent propagation of air pollution (specifically particulate matter and gases) originating from moving emission sources, such as snowmobiles. The study focuses on the unique environmental conditions of Longyearbyen, Spitsbergen, where thermal inversion—a phenomenon where warm air traps dense, cold air near the ground—significantly exacerbates pollutant accumulation.

Mathematically, the problem is modeled using a time-dependent advection-diffusion equation:
$\frac{\partial u}{\partial t} - \nabla \cdot (\epsilon \nabla u) + \mathbf{b} \cdot \nabla u = f(\mathbf{x}, t)$
where $u$ is the pollutant concentration, $\epsilon$ is the diffusion coefficient, $\mathbf{b}$ is the advection field (air movement), and $f$ is the source term.

2. Methodology

The authors propose a novel computational framework called Collocation-based Robust Variational Physics Informed Neural Networks (CRVPINN). The methodology evolves through several stages:

Standard PINNs: Traditional Physics-Informed Neural Networks minimize the strong-form PDE residual. However, the authors note that a small residual loss does not guarantee a small numerical error in the solution.
Variational PINNs (VPINNs): To improve stability, the authors transform the strong-form residuals into a weak-form residual by integrating the equation against a set of test functions. This incorporates initial and boundary conditions directly into the loss function.
Robust VPINNs (RVPINN): To ensure the loss function is directly related to the true approximation error, the authors introduce a robust loss function using the inverse of a Gram matrix ( $G^{-1}$ ). This ensures the loss is spectrally equivalent to the error norm.
CRVPINN (The Core Innovation): To overcome the high computational cost of continuous 3D integration required by VPINNs, the authors introduce a collocation-based strategy. They discretize the space-time domain and replace continuous integrals with discrete inner products and finite-difference gradients. This allows the framework to benefit from the mathematical robustness of variational methods while maintaining the speed of collocation-based training.

3. Key Contributions

Mathematical Framework: The first introduction of a time-dependent CRVPINN method for non-stationary advection-diffusion problems.
Error Bound Guarantee: Proving that the Riesz residual norm is spectrally equivalent to the true error norm ( $\|u_\theta - u^*\|_U$ ), meaning the training loss serves as a reliable estimate of numerical accuracy.
Computational Efficiency: Replacing expensive integration with a discrete Gram matrix and LU factorization to accelerate training.
Empirical Validation: Combining advanced machine learning simulations with real-world, in-field sensor measurements to validate the model's relevance to actual environmental hazards.

4. Results

Numerical Convergence: The CRVPINN method demonstrated superior reliability. While standard PINN loss might reach $10^{-6}$ without guaranteeing accuracy, the CRVPINN loss converged to $0.005$, providing a mathematically bounded error estimate of approximately $0.07$ in the $H^1_0$ norm.
Simulation Findings: The simulations confirmed that thermal inversion in Spitsbergen prevents pollutants from dispersing vertically, trapping snowmobile emissions near the ground and increasing local concentrations.
Field Measurements: Using Airly sensors on moving snowmobiles, the researchers recorded significant spikes in $NO_2$ , $PM_{2.5}$ , and $PM_{10}$ (reaching up to $250 \mu g/m^3$ during acceleration). Stationary measurements also showed $NO_2$ and $PM_{2.5}$ levels exceeding EU safety standards and targets, corroborating the simulation's prediction of localized pollution accumulation.

5. Significance

This work bridges the gap between high-level mathematical theory (variational stability and error analysis) and practical environmental engineering. By providing a robust, fast, and error-aware simulation tool, the CRVPINN framework can be used by policymakers and environmental scientists to predict air quality degradation in complex topographies. The ability to model moving sources in time-dependent scenarios makes it highly applicable to urban traffic pollution and industrial emission monitoring.