Abstract:Air pollution significantly threatens human health and ecosystems, necessitating effective air quality prediction to inform public policy. Traditional approaches are generally categorized into physics-based and data-driven models. Physics-based models usually struggle with high computational demands and closed-system assumptions, while data-driven models may overlook essential physical dynamics, confusing the capturing of spatiotemporal correlations. Although some physics-informed approaches combine the strengths of both models, they often face a mismatch between explicit physical equations and implicit learned representations. To address these challenges, we propose Air-DualODE, a novel physics-informed approach that integrates dual branches of Neural ODEs for air quality prediction. The first branch applies open-system physical equations to capture spatiotemporal dependencies for learning physics dynamics, while the second branch identifies the dependencies not addressed by the first in a fully data-driven way. These dual representations are temporally aligned and fused to enhance prediction accuracy. Our experimental results demonstrate that Air-DualODE achieves state-of-the-art performance in predicting pollutant concentrations across various spatial scales, thereby offering a promising solution for real-world air quality challenges.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve the key challenges in air pollution prediction, especially for the complex dynamics in open systems. Specifically, the authors propose the following core issues: 1. **Limitations of existing methods**: - **Physics - based methods**: These methods usually rely on solving ordinary differential equations (ODEs) or partial differential equations (PDEs). Although they can accurately capture the diffusion and convection processes of pollutants, they are computationally expensive and assume that the system is closed, that is, the total mass remains constant. However, in reality, air pollution systems are open, and pollutants constantly enter and leave the system boundaries. - **Data - driven methods**: These methods are good at extracting spatio - temporal correlations from historical data, but lack the integration of physical principles, which may lead to incomplete or incorrect representations of spatio - temporal relationships. 2. **Mismatch between physical equations and deep - learning representations**: - Physical equations provide an explicit and interpretable framework for modeling changes in pollutant concentrations, and each variable has a clear physical meaning. Deep - learning methods, on the other hand, capture spatio - temporal dependencies through implicit representations and lack direct physical interpretations. This mismatch makes it difficult to incorporate physical knowledge into data - driven models. 3. **Complexity of open systems**: - In open systems, pollutants not only spread within the region but also enter and leave the system through the boundaries. In addition, industrial activities and vehicle emissions introduce new pollutants, while natural elements such as forests and lakes absorb pollutants. Therefore, the traditional closed - system assumption cannot accurately describe the physical phenomena in open systems. To solve these problems, the authors propose a new method called Air - DualODE, which combines dual - branched neural ODEs (Neural ODEs) to predict air quality. Specifically: - **First branch**: Apply the physical equations of open systems to capture spatio - temporal dependencies and learn physical dynamics. - **Second branch**: Identify dependencies not covered by the first branch in a completely data - driven manner. - **Fusion module**: Align and fuse the time of these two branches to improve prediction accuracy. Through this method, Air - DualODE can more accurately simulate the spread of air pollution in open systems, thus providing a better air quality prediction solution. Experimental results show that Air - DualODE achieves state - of - the - art performance in predicting pollutant concentrations at different spatial scales. ### Formula summary - **Continuity equation**: \[ \frac{\partial X}{\partial t}+\nabla\cdot(X\vec{F}) = 0 \] where \(X\) is the air pollutant concentration, \(\vec{F}\) is the particle flux, and \(\nabla\cdot\) is the divergence operator. - **Diffusion - convection equation**: \[ \frac{\partial X}{\partial t}=k\cdot\nabla^{2}X-\nabla\cdot(X\cdot\vec{v}) \] where \(k\) is the diffusion coefficient, \(\vec{v}\) is the wind field, and \(\nabla^{2}\) is the Laplace operator. - **Discrete boundary - aware diffusion - convection equation (BA - DAE)**: \[ F_{P}(X_{t};\Theta)=\frac{dX}{dt}=\alpha\odot(-k\cdot L_{\text{diff}}X)+(1 - \alpha)\odot(L_{\text{adv}}X)+\beta X \] where \(\Theta\) is the learning parameter for approximating the graph Laplacian operator, \(\alpha\in\mathbb{R}^{N\times1}\) is a gating value, and \(\beta\) is the correction term coefficient. Through these formulas and methods, Air - DualODE can better handle the air pollution prediction problems in open systems.

Air Quality Prediction with Physics-Informed Dual Neural ODEs in Open Systems