pbdl-book

Why we included this project

Most machine learning courses treat physics as just another dataset, which is why the Physics-based Deep Learning book stands out. It's an interactive Jupyter book where every chapter pairs a concept with runnable notebooks, so you can move from reading about a technique to running it yourself in the same sitting. The content is practical rather than theoretical: physical loss constraints, differentiable simulation, and diffusion-based probabilistic models that act as faster stand-ins for conventional solvers. There's also material on leaning on a full simulator during training for inverse problems, and on how architecture choices, like continuous versus discrete representations or structured versus unstructured graph meshes, change what works. It deliberately skips the broad survey and assumes you already know the basics of deep learning and numerical simulation, which makes it a useful on-ramp for applied researchers and engineers moving into learned surrogates for fluid dynamics and PDE problems.