Course Contents

Course Contents#

Instructor: Carlos D. Hoyos
Institution: Universidad Nacional de Colombia, Facultad de Minas, Sede Medellín

Course Description#

This graduate-level course explores the intersection of AI, ML, and atmospheric sciences, emphasizing practical skills in data analysis, modeling, and prediction. Topics range from foundational ML principles to specialized applications like extreme weather forecasting and climate modeling.

Syllabus#

Weeks 1-3: Introduction to AI, ML, and Core Concepts in Atmospheric Sciences#

Week 1: Overview of AI/ML in Atmospheric Sciences#

  • Topics: Introduction to AI and ML concepts in the context of atmospheric sciences, history, and evolution of AI applications in this field, from early statistical models to deep learning and neural networks.

  • Readings: Review papers on ML applications in atmospheric sciences, societal and environmental impacts.

Week 2: Fundamental Concepts in Atmospheric Sciences#

  • Topics:

    • Forces in the Atmosphere: Understanding gravitational, Coriolis, and frictional forces, as well as their effects on atmospheric motion.

    • Navier-Stokes Equations (NS): Introduction to the NS equations as the foundation for describing fluid motion in the atmosphere. Discuss the challenges of solving these equations and approximations used in atmospheric models.

    • Basic Atmospheric Thermodynamics: Key concepts, including temperature, pressure, density, and their relationships, that govern atmospheric behaviors.

    • Traditional NWP Models: Overview of Numerical Weather Prediction (NWP), model grids, parameterizations, and how these models integrate physical laws to simulate the atmosphere.

  • Readings: Atmospheric Science: An Introductory Survey by John M. Wallace and Peter V. Hobbs (selected chapters on atmospheric dynamics and thermodynamics).

Week 3: Atmospheric Modeling, Forecasting, and Modern Challenges#

  • Topics:

    • Forecasting Techniques: Comparison of traditional and modern forecasting approaches, including statistical, dynamical, and AI-based models.

    • Challenges in Atmospheric Sciences: Discuss the limitations in current NWP models, including data sparsity, computational costs, model biases, and the need for downscaling to improve regional forecasts.

    • Societal and Environmental Needs: Explore how improved atmospheric models address societal needs, including disaster preparedness, climate resilience, air quality, and water resource management.

    • Current State-of-the-Art: Introduction to emerging AI/ML applications such as data-driven weather prediction, hybrid physics-AI models, and the role of big data in advancing forecasting skill.

  • Readings: Recent journal articles on state-of-the-art challenges in weather and climate prediction, such as downscaling techniques, data assimilation, and the integration of AI into traditional NWP systems.

Weeks 4-5: Fundamentals of Machine Learning#

  • Topics: Supervised vs. unsupervised learning, key algorithms, model evaluation, and metrics. Introduction to regression and classification techniques applied to atmospheric data.

  • Readings: Machine Learning: A Probabilistic Perspective by Kevin Murphy.

Weeks 6-7: Neural Networks and Deep Learning Foundations#

  • Topics: Basics of neural networks, convolutional neural networks (CNNs), and recurrent neural networks (RNNs). Introduction to applications of CNNs and RNNs in recognizing atmospheric patterns, e.g., cloud formation and precipitation.

  • Readings: NVIDIA research on deep learning for weather prediction.

Weeks 8-9: Advanced Neural Networks for Atmospheric Applications#

  • Topics: Transformers, Long Short-Term Memory (LSTM) networks, and sequence-to-sequence models. Emphasis on ML-based weather prediction research from leading companies like DeepMind, Microsoft, and NVIDIA, exploring their techniques and recent advancements.

  • Readings: DeepMind and Microsoft papers on transformers for temporal forecasting.

Weeks 10-11: Generative Models and Uncertainty Quantification#

  • Topics: Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and methods for quantifying uncertainty in predictions. Applications in synthetic data generation for extreme weather scenarios.

  • Readings: Research by Huawei on generative models for atmospheric applications.

Weeks 12-13: Physics-Informed AI and Hybrid Modeling#

  • Topics: Physics-informed neural networks (PINNs) and hybrid approaches combining ML with numerical weather prediction (NWP). Exploration of how integrating physical constraints into ML models improves prediction accuracy and reliability in atmospheric contexts.

  • Readings: ECMWF MOOC materials and NVIDIA papers on physics-informed models.

Weeks 14-15: Interpretability and Dimensionality Reduction#

  • Topics: Explainable AI techniques (e.g., SHAP values, LIME) and dimensionality reduction methods (e.g., UMAP, t-SNE) for visualizing high-dimensional climate data. Discussion on the importance of interpretability in ensuring model transparency and understanding in climate and weather forecasting.

  • Readings: Recent research on interpretability in ML for atmospheric sciences.

Weeks 16: Climate Modeling, Downscaling, and Advanced Applications#

  • Topics: High-resolution climate modeling, downscaling techniques, and transfer learning applications in climate science. Focus on leveraging transfer learning to enhance predictive capability across spatial and temporal scales in climate models.

  • Readings: Microsoft and DeepMind research on transfer learning in climate science.

Learning Outcomes#

By the end of this course, students will:

  1. Understand fundamental AI and ML techniques and their applications in atmospheric sciences.

  2. Gain hands-on experience with neural networks, and interpretable ML.

  3. Apply physics-informed AI and downscaling techniques to atmospheric datasets.

  4. Develop skills in working with atmospheric data to address challenges in climate resilience, energy forecasting, and environmental management.

Course Resources#

  • Textbooks:

    • Deep Learning by Ian Goodfellow.

    • Machine Learning: A Probabilistic Perspective by Kevin Murphy.

  • Online Materials:

    • ECMWF MOOC, GitHub resources, and recent ML research papers.

  • Tools: Jupyter Notebooks, TensorFlow, Keras, PyTorch.

Assessment#

  • Weekly Assignments: 50% (hands-on coding and problem sets)

  • Final Project: 50% (develop an AI/ML model applied to a real-world atmospheric problem)

Additional Notes#

  • Collaboration: Students are encouraged to participate in discussions, fostering collaborative learning and continuous improvement.

Contents