제목: Deep Learning Application for emulator, air quality forecasting, estimating surface air quality, and wildfire forecasting​

​​

​​​​​​​​​​​​​​​​​​연사: 최윤수 교수 (University of Houston)​

​

장소: 과학관 B103​호​​

Abstract:

Accurate and efficient air quality and wildfire risk forecasting are essential for exposure assessment, policy-making, and early warning systems. This study presents a suite of deep learning-based approaches to emulate chemical transport models (CTMs), forecast air quality in real time, estimate surface pollutant concentrations using limited surface measurements or remote sensing data, and predict high-resolution fire weather indices (FWI). First, a U-Net-based emulator was developed to replicate CMAQ outputs of NO₂, O₃, and PM₂.₅ using the EQUATES dataset (2015–2019), achieving over 1,000×computational speed-up while maintaining high fidelity. Second, for real-time air quality forecasting in South Korea, an attention-based graph neural network (AGATNet) and a physics-informed Transport-informed GNN (TiGNN) outperformed CNN approaches, improving extreme event capture and long- horizon forecasts. Third, machine learning models (Random Forest, PCNN-DNN, and Deep-CNN) estimated surface PM₂.₅, NO₂, and MDA8 ozone at high spatiotemporal resolutions using satellite, meteorological, and land-use data, achieving superior accuracy compared to conventional methods and enabling health impact assessments. Finally, hybrid GNN-based models(GNN-TCNN, GNN-LSTM, GNN-DeepAR) were applied for high-resolution FWI forecasting, with GNN-TCNN achieving the best long-term performance. Collectively, these deep learning frameworks substantially enhance computational efficiency and predictive accuracy, supporting scalable, policy-relevant applications in air quality management and wildfire risk mitigation.​​