Register now After registration you will be able to apply for this opportunity online.
Predict and understand atmospheric radiation flow with graph neural networks
Graph neural networks (GNNs) have attracted intense interest recently, and so does their explainability. AI can help solve major challenges in environmental sciences, one of them being radiation level forecast and understanding. While many methods using ML techniques and deep neural networks are already used for climate sciences, the power of graph neural networks is still under estimated in climate applications. Radiation is a key component of the Earth’s energy cycle. Convergence of radiative flux is a heat source and an essential component of the numerical solver of the thermodynamic equation in a climate model. The current problem that scientists face is the problem of discontinuity of predicted radiation levels due to the space discretization. The radiative flux obtained from deep learning methods is often not smooth enough to obtain realistic heating rates. Currently, the heating rates are obtained from the radiative flux by a simple vertical finite-difference differentiation to estimate the vertical flux convergence. GNNs are promising tools to predict continuous radiation flow in the atmosphere and from that heating rates. A particular advantage of ML-based methods is their resource efficiency, which has the potential to considerably reduce the compute resources needed for a complete climate integration and allow for high model resolution that is needed to accurately simulate high-impact weather.
Keywords: Radiative flux, graph neural network, explainable AI
**Your Role**
This project will require you to model the atmospheric radiation flow as a graph and explore different GNN architectures to predict continuous radiation levels. This will be done under the constraint of physical equations and the conservation of energy. The radiative flux is considered to be the edges of the graph and the heating rates its vertices. Then, you will investigate explanations produced by post-hoc explainability methods that will confirm the validity of your results. The project will give you hands-on experience with graph neural networks and popular explainability methods (GradInput, Integrated Gradients, GNNExplainer), as well as getting familiar with the reasons behind global warming.
**Requirements**
We are particularly interested in students with a background and research interests in the following areas: deep learning, graph neural networks, interpretability. The student is expected to have good experience with programming in Python and basic knowledge in graph neural networks. We expect our students to be highly motivated to work on pluri-disciplinary topics and to cooperate with their supervisors regularly to discuss current progress and next steps. We also expect curious students with interest in solving climate challenges.
**What we offer**
This topic will give you the opportunity to contribute to solving environmental challenges like global warming. With this project, you will also learn and get involved in current research problems in two hot topics: explainability and GNNs. After this Master thesis, you will be familiar with the field of GNNs and how they can be applied in concrete and meaningful ways. We also offer you an interdisciplinary work environment and the possibility to connect with people in related research domains. You will have access to training and climate model data at the Institute for Atmospheric and Climate Science and ETH’s high-performance computer Euler including CPU nodes and a GPU node (Tesla A-100 80GB).
**Your Role**
This project will require you to model the atmospheric radiation flow as a graph and explore different GNN architectures to predict continuous radiation levels. This will be done under the constraint of physical equations and the conservation of energy. The radiative flux is considered to be the edges of the graph and the heating rates its vertices. Then, you will investigate explanations produced by post-hoc explainability methods that will confirm the validity of your results. The project will give you hands-on experience with graph neural networks and popular explainability methods (GradInput, Integrated Gradients, GNNExplainer), as well as getting familiar with the reasons behind global warming.
**Requirements**
We are particularly interested in students with a background and research interests in the following areas: deep learning, graph neural networks, interpretability. The student is expected to have good experience with programming in Python and basic knowledge in graph neural networks. We expect our students to be highly motivated to work on pluri-disciplinary topics and to cooperate with their supervisors regularly to discuss current progress and next steps. We also expect curious students with interest in solving climate challenges.
**What we offer**
This topic will give you the opportunity to contribute to solving environmental challenges like global warming. With this project, you will also learn and get involved in current research problems in two hot topics: explainability and GNNs. After this Master thesis, you will be familiar with the field of GNNs and how they can be applied in concrete and meaningful ways. We also offer you an interdisciplinary work environment and the possibility to connect with people in related research domains. You will have access to training and climate model data at the Institute for Atmospheric and Climate Science and ETH’s high-performance computer Euler including CPU nodes and a GPU node (Tesla A-100 80GB).
The goal of this project is to model the atmospheric radiation flow in an atmospheric column as a graph and apply graph neural networks to predict radiation while avoiding the discontinuity problem of current methods. Training data is obtained from state-of-the-art climate models and its physics-based radiation flux solver. In addition, we propose to explain the radiation flow using post-hoc explainability methods for GNNs and due to an ongoing collaboration with the Swiss Data Science Center, the GNN-based result may be compared to the results obtained from RNNs, CNNs, and a Random Forest.
The goal of this project is to model the atmospheric radiation flow in an atmospheric column as a graph and apply graph neural networks to predict radiation while avoiding the discontinuity problem of current methods. Training data is obtained from state-of-the-art climate models and its physics-based radiation flux solver. In addition, we propose to explain the radiation flow using post-hoc explainability methods for GNNs and due to an ongoing collaboration with the Swiss Data Science Center, the GNN-based result may be compared to the results obtained from RNNs, CNNs, and a Random Forest.
Interested students should send their CV, academic transcripts and explain why they are interested to Kenza Amara (kenza.amara@ai.ethz.ch). The proposed Master Thesis will be supervised by Prof. Ce Zhang in collaboration with Prof. Sebastian Schemm.
Interested students should send their CV, academic transcripts and explain why they are interested to Kenza Amara (kenza.amara@ai.ethz.ch). The proposed Master Thesis will be supervised by Prof. Ce Zhang in collaboration with Prof. Sebastian Schemm.
Master_Thesis_Radiation.pdf
