Research map / Probabilistic and scientific modeling / Molecules and proteins
AirPhyNet: Harnessing Physics-Guided Neural Networks for Air Quality Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AirPhyNet, a physics-guided neural ODE network that encodes diffusion and advection dynamics on a sensor graph to forecast PM2.5 concentrations up to 72 hours ahead.
Problem
Long-horizon PM2.5 air quality forecasting is inaccurate in sparse-data settings and deep models are black boxes lacking physical grounding and interpretability.
Approach
Encodes diffusion (distance-based graph Laplacian) and advection (wind-derived flow-field graph Laplacian) as a graph differential equation, fused with a learned gate, and integrated into a GRU encoder + neural ODE solver + decoder architecture.
Key results
- On Beijing 24h/48h/72h, AirPhyNet reports MAE 29.11/36.69/42.23 versus the best baseline AirFormer's 29.62/38.43/43.39 (and STGCN/GTS in some columns).
- In the sparse-data setting (3:1:6 split), AirPhyNet reduces MAE and RMSE by 8.3% and 8.1% on average versus the second-best method.
- Sudden-change prediction: AirPhyNet achieves average reductions of 6.2% MAE and 4.6% RMSE versus AirFormer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine diffusion and advection processes (ablation) | AirPhyNet · Beijing and Shenzhen, 24h–72h | MAE | Helped |
| combine diffusion and advection processes (ablation) | AirPhyNet · Beijing and Shenzhen, 24h–72h | RMSE | Helped |
| use diffusion only, excluding advection (ablation) | AirPhyNet · Beijing PM2.5 prediction | prediction performance | Hurt |
| use advection only, excluding diffusion (ablation) | AirPhyNet · Beijing PM2.5 prediction | prediction performance | Hurt |
| replace GRU encoder with TCN encoder (ablation) | AirPhyNet · Beijing PM2.5 prediction | prediction performance | Hurt |
| use AirPhyNet instead of the strongest competing baseline | AirPhyNet · Beijing and Shenzhen, 24h–72h | MAE: 3.7% | Helped |
| use AirPhyNet instead of the strongest competing baseline | AirPhyNet · Beijing and Shenzhen, 24h–72h | RMSE: 6.1% | Helped |
| use AirPhyNet instead of the second-best method | AirPhyNet · Sparse data, Beijing and Shenzhen | MAE: 8.3% | Helped |
| use AirPhyNet instead of the second-best method | AirPhyNet · Sparse data, Beijing and Shenzhen | RMSE: 8.1% | Helped |
| use AirPhyNet instead of AirFormer | AirPhyNet · Sudden-change prediction, Beijing and Shenzhen | MAE: 6.2% | Helped |
| use AirPhyNet instead of AirFormer | AirPhyNet · Sudden-change prediction, Beijing and Shenzhen | RMSE: 4.6% | Helped |
| add wind-related attributes to baseline models | Baseline models other than PM2.5-GNN · Air quality prediction | prediction performance | Hurt |
Surprising findings
- Most baselines degraded when wind attributes were added as inputs, so they were reported without wind features (only PM2.5-GNN used them).
Limitations
- Focused only on PM2.5; not validated on pollutants dominated by chemical reactions.
- Uses only two datasets from China, both with few stations (11 and 35).
- Improvements over the best baseline are modest on several horizons (e.g., 1.7% to 4.5% MAE on Beijing).
Experimental setup
Datasets: Beijing air quality dataset (Harvard Dataverse) Shenzhen urban air dataset (Microsoft Research)
Benchmarks: Beijing PM2.5 dataset Shenzhen PM2.5 dataset
Metrics: MAE RMSE
Baselines: HA VAR DCRNN STGCN GMAN GTS PM2.5-GNN AirFormer Latent-ODE ODE-LSTM
Builds on: Neural ODE Latent ODE DCRNN AirFormer GRU
Closest papers by meaning
- Air Quality Prediction with Physics-Guided Dual Neural ODEs in Open Systems (ICLR 2025 · similarity 0.93)
- AirRadar: Inferring Nationwide Air Quality in China with Deep Neural Networks (AAAI 2025 · similarity 0.87)
- SNN-PDE: Learning Dynamic PDEs from Data with Simplicial Neural Networks (AAAI 2024 · similarity 0.84)
- CoPINN: Cognitive Physics-Informed Neural Networks (ICML 2025 · similarity 0.82)
- Physics-Guided Foundation Model for Scientific Discovery: An Application to Aquatic Science (AAAI 2025 · similarity 0.81)
- PINNacle: A Comprehensive Benchmark of Physics-Informed Neural Networks for Solving PDEs (NeurIPS 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.