Part III

Discovery Through Data and Models

Machine learning methods that extract structure, generate hypotheses, and uncover hidden patterns

Overview

Part III covers the data-driven and model-driven techniques at the heart of computational discovery. From exploratory analysis and representation learning through foundation models and reasoning architectures, these chapters show how modern ML finds anomalies, infers causal structure, and generates novel hypotheses. The part also addresses Bayesian methods, physics-informed scientific ML, generative models, and symbolic regression for recovering interpretable laws from data.

Chapters

Chapter 25 Exploratory Discovery Unsupervised exploration, clustering, dimensionality reduction, and data-driven hypothesis generation. Chapter 26 Representation Learning Learning meaningful embeddings and latent spaces that capture the structure of scientific data. Chapter 27 Scientific Foundation Models Large pretrained models for proteins, molecules, climate, and other scientific domains. Chapter 28 Multimodal Scientific AI Integrating text, images, spectra, sequences, and graphs for cross-modal scientific reasoning. Chapter 29 Reasoning Models Chain-of-thought, tree search, and neuro-symbolic architectures for structured scientific reasoning. Chapter 30 Anomaly and Novelty Detection Identifying unexpected patterns, outliers, and novel phenomena that signal potential discoveries. Chapter 31 Causal Discovery Learning causal graphs from observational and interventional data to explain mechanisms. Chapter 32 Bayesian Discovery Probabilistic modeling, Bayesian optimization, and uncertainty-aware discovery workflows. Chapter 33 Scientific Machine Learning Physics-informed neural networks, neural operators, and hybrid models that respect physical laws. Chapter 34 Generative Models for Discovery VAEs, diffusion models, and flow-based architectures for generating molecules, materials, and designs. Chapter 35 Symbolic Regression Recovering interpretable mathematical expressions and physical laws directly from data.