重点：CEREA 实验室不属于ZRR 范畴，无需接受ZRR 审查
导师介绍：Sibo Cheng，目前在伦敦帝国理工大学计算机系任博士后研究员，将于2024年4月入职巴黎高科路桥大学 (Ecole des ponts paristech）担任准聘教授 （chaire de professeur junior），独立PI。研究方向为机器学习，数据同化及其在高维动态问题（特别是环境物理场）中的应用，在计算科学和人工智能顶刊顶会 JCP, CMAME, JSC, CEJ, TNNLS, Neurips 等刊物发表多篇论文。
要求：课题组将招收2名或以上全奖硕士实习生（stage fin d'étude）（5-6个月）和一名全奖 gap year research assistant （césure）.
1. 硕士实习生和Research assistant必须为欧盟（不包含英国）大学/工程师学校在读.
2. 就读相关专业，对科研工作有兴趣和热情
3. 熟练掌握python 编程，有pytorch项目经验优先。
欢迎感兴趣的同学联系并附上英语或法语简历至 sibo.cheng@imperial.ac.uk
课题组后续将招聘全奖博士和博士后（无国籍，教育地区限制，无审核要求），欢迎感兴趣的同学联系.
项目介绍：
Project 1：Utilizing Multi-Scale Graph Neural Networks for Air Pollution Prediction in Major Cities
Background: Air pollution in large urban areas is a critical environmental issue, with complex dynamics influenced by various factors such as traffic, industrial activities, and weather conditions. Traditional methods (eg CNN or RNN based) for predicting air pollution levels often struggle with the unstructured nature and complex geometry of data involved in the spread of pollutants like CO2.
Objective: The primary objective of this project is to develop an efficient and accurate predictive model for air pollution in big cities, specifically focusing on CO2 levels. This will be achieved by leveraging the capabilities of multi-scale graph neural networks (GNNs) to handle the unstructured data and complex geometric relationships inherent in urban air pollution scenarios. Depending on the final results, this student will have chance to disserminate the results in major computer science/computing conference or journals.
Reference：https://arxiv.org/abs/2302.06186
Project 2：Using Transformers (Masked Autoencoder) for Predicting high-dimensional Dynamical Systems with Irregular Time Steps
Background: The accurate prediction of dynamical systems, especially those with irregular time steps and sparse observations, is a significant challenge in various fields like climate prediction, wildfire management, and fluid dynamics. Traditional models, such as CNNs or RNNs, often struggle with irregular temporal data and high-dimensional systems. This project proposes the development of a transformer-based model to address these challenges.
Objective: The primary objective of this project is to build and evaluate transformer models capable of predicting high-dimensional dynamical systems using sparse and irregular time observations. The project aims to leverage the unique strengths of transformers in handling irregular time steps and missing observations, which are limitations in traditional RNN or CNN models. Additionally, the project will explore the integration of these transformer models with reduced-order models, like autoencoders, to manage the complexity of high-dimensional systems effectively.
Data：The student will work on both simulated data (e.g. fluid dynamics) and climate/ocean data from remote sensing
Key references: https://arxiv.org/abs/2301.08871
