School of Science Department of Mathematics 51 Use Machine Learning Technique to Retrieve Air Quality Data of SO2 Supervisor: FUNG Jimmy Chi Hung / MATH Co-supervisor: LU Xingcheng / ENVR Student: SUN Haochen / DSCT Course: UROP2100, Fall In this report, we will explore multiple deep learning model structures based on the Long Short-term Memory (LSTM) layers that predict the concentration of FSPMC and O3 in the air. With ground observation data of weather and air quality for the past 72 hours as the input to a baseline model, we try to predict the FSPMC and O3 concentration of the future 48 hours. We modify the model by adding WRF-CMAQ prediction of the target hours as an additional input, applying 1D-CNN layers to the input, and applying non-trivial lastlayer activation functions and loss functions. With the methods listed above, we achieve a relatively high accuracy in the task. Use Machine Learning Technique to Retrieve Air Quality Data of SO2 Supervisor: FUNG Jimmy Chi Hung / MATH Co-supervisor: LU Xingcheng / ENVR Student: SHA Yu Hin / COSC Course: UROP3100, Fall Air pollution has been a crucial problem for its adverse health and economic impact, hence the need for precise air quality forecasting. With advancements in machine learning, neural networks have been a dominant tool for air quality forecasting due to their high time series prediction accuracy. Models, including recurrent neural networks and Seq-2-Seq networks, have been thoroughly experimented with in the field. A novel network architecture called Transformer has been proposed for performing machine translation. In this project, we adopted this Transformer approach for making 24-to-24 hours forecasting in PM2.5 and O3 forecast and assessed its performance through 53 air monitoring stations across China. Use Machine Learning Technique to Retrieve Air Quality Data of SO2 Supervisor: FUNG Jimmy Chi Hung / MATH Co-supervisor: LU Xingcheng / ENVR Student: YI Sophia / COMP Course: UROP2100, Spring Fine inhalable particles, with diameters that are generally 2.5 micrometers and smaller ( 2.5 ) are the main component of air pollutants emitted to the atmosphere. They are the major cause of reduced visibility. Moreover, exposure to such practices poses a threat to health. Therefore, monitoring the concentrations of 2.5 is crucial for air quality management. Obtaining accurate estimates of 2.5 from the ground monitoring sensors is a challenging task due to the limitation on the resolution of measured data. So, machine learning is taken to compensate for such disadvantages. The previous approach of using a backpropagation network for prediction achieved relatively good performance, but the performance could be further improved. A transfer learning approach based on the convolutional network is taken in this report. This project builds up the model to estimate 2.5 from 2015 to 2017 using 15 parameters including Aerooptical Depth. By training these datasets, the model may reflect the concentration of 2.5 in China.