ABOUT EVENT

This workshop will introduce the concept of physics-informed machine learning and explore why physical constraints matter in AI models for environmental science. Using a hands-on example from atmospheric science, participants will train a simple neural network – first as a purely data-driven model, then with physics-informed constraints. By comparing the two approaches, participants will gain an intuition for when and why physical constraints can improve model behavior, particularly in terms of interpretability and generalization. Participants will work through a guided Jupyter notebook in which they build, train, and evaluate two versions of a simple neural network: one purely data-driven and one incorporating physical constraints. They will compare model outputs, visualize predictions, and discuss the trade-offs between flexibility and physical consistency.

Prerequisites

• Basic Python programming and familiarity with Jupyter notebooks. Prior knowledge of machine learning helpful, but not required. No prior knowledge of atmospheric science needed.

Learning Objectives

By the end of this workshop, the goal is for attendees to be able to:

• Explain the difference between purely data-driven and physics-informed machine learning models
• Implement a simple physics-informed neural network in Python
• Evaluate and compare model outputs across both approaches
• Identify scenarios in environmental data science where physics-informed approaches are beneficial

Tools Used

• Python, Jupyter notebooks, NumPy, Matplotlib, PyTorch (or TensorFlow)