Introduction¶
Part IV moves from building models that predict well to building models that can be trusted: understanding why a model makes the predictions it does, quantifying how confident those predictions are, and blending machine learning with physical knowledge rather than treating it as a black box.
Chapter 9 covers explainable AI (XAI) — partial dependence plots, permutation feature importance, and SHAP-based explanations, applied first to tabular classifiers (Titanic survival, wine quality) and then to a convolutional neural network trained on MNIST.
Chapter 10 covers generative modeling and uncertainty quantification — distributional regression with the continuous ranked probability score, then autoencoders, GANs, and a diffusion model, all trained on CIFAR-10.
Chapter 11 covers hybrid modeling — combining a physics-based numerical model with a trained CNN emulator, applied to glacier ice-flow simulation.
As in Parts II and III, each chapter pairs one tutorial notebook with separate exercise notebooks, rather than Part I’s one-lecture-one-exercises structure.