This book is organized as a single arc in seven parts: foundations, software engineering, data and models, knowledge, simulation, domain science, and autonomous discovery. Fifty-eight chapters and seven appendices take you from "what does it mean to discover something?" to "a self-driving laboratory that designs, runs, and interprets its own experiments." A hands-on Discovery Workbench platform threads through every part, so each idea is something you build and run, not merely read about. The full map lives in the Table of Contents; this page walks the arc so you know where each part fits and why the order matters.
Part I: Foundations of Discovery AI (Chapters 1 to 6)
Part I establishes what discovery is and how machines can pursue it. Chapter 1 frames discovery as search through a space of hypotheses, experiments, and models, drawing on the philosophy of science from Popper's falsification to modern computational epistemology. Chapter 2 dissects the scientific method as an algorithmic loop: observation, hypothesis generation, experimental test, revision. Chapters 3 and 4 supply the representational and reasoning machinery that discovery systems need: ontologies, knowledge graphs, first-order logic, probabilistic inference, and constraint satisfaction. Chapter 5 surveys the three complementary routes to discovery (data-driven, model-driven, and simulation-driven) and explains when each dominates. Chapter 6 introduces the Discovery Workbench, the book's reference architecture that unifies these routes under a single platform, with agents, registries, and provenance tracking wired in from the start.
Nothing in Part I is merely philosophical. The vocabulary of search spaces, hypothesis states, and knowledge representations recurs verbatim in every system the later parts build.
Part II: Discovery Through Software Engineering and Vibe Coding (Chapters 7 to 24)
Part II is the largest part of the book, because building discovery systems is, at bottom, a software engineering problem, and AI has transformed every stage of that engineering. It begins by recasting software development itself as a discovery process (Chapter 7) and mapping the landscape of AI-assisted software engineering, from autocompletion to autonomous coding agents (Chapter 8). Chapters 9 through 12 define and explore vibe coding as a four-phase discipline: specification through natural language and intent capture, steering through iterative prompt refinement, verification through test generation and formal checks, and repair through automated debugging and self-healing pipelines.
From there the part deepens. Chapter 13 covers the transition from prompting to programming: when a conversation with an LLM becomes a maintained codebase. Chapter 14 introduces context engineering, the art of feeding large models the right code, documentation, and domain knowledge at the right moment. Chapter 15 builds MCP (Model Context Protocol) servers tailored to scientific workflows, connecting coding agents to lab instruments, data repositories, and simulation engines. Chapters 16 through 18 apply AI-assisted discovery to the three classic software unknowns: requirements, architectures, and algorithms. Chapter 19 addresses AI-assisted implementation at scale, and Chapter 20 assembles multi-agent software teams where planner, coder, reviewer, and tester agents collaborate on a shared repository.
The closing stretch covers the operational side: AI-assisted testing and debugging (Chapter 21), security analysis and DevOps automation (Chapter 22), the MLOps/LLMOps/AgentOps stack for deploying and monitoring AI-native systems (Chapter 23), and a rigorous evaluation framework for coding agents, from pass@k benchmarks to real-world productivity studies (Chapter 24). The part concludes with autonomous software organizations, where entire development pipelines run with minimal human oversight.
Part III: Discovery Through Data and Models (Chapters 25 to 35)
Part III turns to the models that extract patterns, generate hypotheses, and propose new structures directly from scientific data. Chapter 25 opens with exploratory discovery: automated pipelines for anomaly detection, clustering, and dimensionality reduction that surface unexpected structure in high-dimensional datasets. Chapter 26 covers representation learning, from autoencoders and contrastive methods to the embeddings that let a model "see" molecules, genomes, and materials in a shared latent space.
Chapters 27 and 28 form the part's centerpiece: scientific foundation models and multimodal scientific AI. You will work with protein language models (ESM, ProtTrans), molecular representations (SchNet, DimeNet), materials property predictors, and genomic foundation models, then learn to fuse modalities (spectra, images, text, structures) into unified systems. Chapter 29 examines reasoning models that chain logical steps over scientific evidence rather than pattern-matching surface statistics. Chapters 30 through 32 cover three distinct discovery modes: anomaly and novelty detection for finding what a model does not expect, causal discovery for learning directed graphs of influence from observational data, and Bayesian discovery for updating beliefs as evidence accumulates.
The final three chapters address model families built for scientific structure. Chapter 33 covers scientific machine learning: physics-informed neural networks (PINNs), neural operators (FNO, DeepONet), and equivariant networks that respect the symmetries of physical law. Chapter 34 explores generative models for discovery, including diffusion models and flow matching applied to molecular and materials generation. Chapter 35 closes with symbolic regression and equation discovery, where systems like PySR and AI Feynman recover interpretable mathematical laws from data.
Part IV: Discovery Through Knowledge (Chapters 36 to 41)
Part IV addresses the fact that most scientific knowledge lives not in datasets but in text: papers, patents, clinical notes, lab notebooks. Chapter 36 builds literature-mining pipelines that extract entities, relations, and claims from millions of documents using NER, relation extraction, and large language models fine-tuned on scientific corpora. Chapter 37 constructs retrieval-augmented discovery (RAD) systems that ground hypothesis generation in the full text of a scientific literature base, going beyond simple RAG to multi-hop retrieval and evidence synthesis.
Chapter 38 tackles knowledge graph discovery: constructing, completing, and reasoning over large-scale scientific knowledge graphs such as those mapping drug-target interactions or gene-disease associations. Chapter 39 turns those graphs and literature embeddings into hypothesis generators, systems that propose novel connections (a gene linked to a disease, a material suitable for a new application) and rank them by plausibility. Chapter 40 orchestrates research agents, autonomous systems that read papers, design follow-up experiments, and write preliminary reports. Chapter 41 closes the loop with scientific claim validation: automated fact-checking pipelines that assess reproducibility, statistical soundness, and consistency with prior work.
Part V: Discovery Through Simulation and Optimization (Chapters 42 to 47)
Part V covers the computational laboratory where hypotheses are tested before any physical experiment runs. Chapter 42 introduces differentiable programming in JAX, teaching you to write scientific simulations whose every step is automatically differentiable, so gradients flow from a loss function back through an entire physical model. Chapter 43 builds on that foundation with scientific simulation at scale: molecular dynamics, finite-element solvers, and agent-based models accelerated by neural surrogates.
Chapter 44 develops world models, learned simulators that internalize the dynamics of a domain and can predict, plan, and imagine counterfactual scenarios. Chapter 45 surveys optimization for discovery, from Bayesian optimization over expensive black-box functions to evolutionary strategies and reinforcement learning for sequential experiment design. Chapter 46 applies these tools to automated experiment design: systems that choose the next experiment to maximize information gain, minimize cost, or satisfy a multi-objective Pareto frontier. Chapter 47 grounds all of this in reproducibility, covering experiment registries, provenance graphs, and artifact management so that every simulated (and later physical) experiment is traceable end to end.
Part VI: Discovery in Scientific Domains (Chapters 48 to 52)
Part VI applies the tools of Parts I through V to five major scientific domains, each with its own data modalities, constraints, and discovery traditions. Chapter 48 covers biology and medicine: AI-driven drug discovery pipelines from target identification to lead optimization, protein design with structure-prediction models, and genomics workflows that link variants to phenotypes. Chapter 49 turns to chemistry and materials science: molecular generation with graph neural networks and diffusion models, machine-learned force fields that replace expensive quantum calculations, and the autonomous laboratory platforms that synthesize and characterize candidates without human intervention.
Chapter 50 addresses physics and engineering: symbolic regression for recovering governing equations, PINNs for solving and inverting partial differential equations, and neural operators that learn mappings between function spaces for fluid dynamics, electromagnetics, and structural mechanics. Chapter 51 covers climate and earth science: foundation models for weather prediction, downscaling, and extreme-event attribution trained on reanalysis data and satellite imagery. Chapter 52 closes with social and economic systems: agent-based modeling of markets and epidemics, causal inference from observational social data, and the particular challenges of discovery in domains where controlled experiments are rare.
Part VII: Autonomous Discovery Systems (Chapters 53 to 58)
Part VII brings everything together into systems that discover with minimal human guidance. Chapter 53 surveys the first generation of AI scientists: The AI Scientist (v1 and v2), Coscientist, and Google's AI Co-Scientist, analyzing what each system automates, what it still requires from humans, and where it fails. Chapter 54 scales from single agents to multi-agent discovery systems, where specialized agents for literature review, hypothesis generation, experiment execution, and manuscript drafting coordinate through shared memory and structured communication protocols.
Chapter 55 enters the self-driving laboratory: fully automated platforms such as the Emerald Cloud Lab and Artificial Chemist, where robotic hardware, AI planning, and closed-loop optimization run 24/7 discovery campaigns. Chapter 56 confronts evaluation head-on, defining metrics and benchmarks for discovery quality (novelty, validity, significance, efficiency) that go beyond the accuracy scores of conventional ML. Chapter 57 addresses responsible discovery AI: bias in training corpora, dual-use risks in biology and chemistry, intellectual property questions around AI-generated inventions, and governance frameworks for autonomous research. Chapter 58 closes with future directions: open problems in scientific reasoning, the path from narrow discovery agents to general-purpose scientific intelligence, and what a mature discovery ecosystem might look like in the next decade.
Beyond the Chapters
Seven appendices support the main text. Appendix A reviews the mathematical foundations (linear algebra, probability, optimization, information theory) that the chapters draw on. Appendix B is a hands-on Python for Discovery AI guide, covering JAX, PyTorch, and the scientific Python stack. Appendix C provides a vibe coding practitioner guide with prompt libraries, agent configurations, and debugging checklists. Appendix D catalogs the discovery AI tool ecosystem: frameworks, datasets, benchmarks, and pretrained models organized by domain. Appendix E offers a research paper reading guide for efficiently extracting methods, claims, and reproducibility signals from the literature. Appendix F covers reproducibility and artifact management: experiment registries, containerized environments, and artifact provenance for both computational and physical experiments. Appendix G presents five capstone project tracks (software discovery, data-driven discovery, knowledge-driven discovery, simulation-driven discovery, and autonomous discovery) that each synthesize multiple parts of the book into an end-to-end system you build, evaluate, and document.
One thread runs through everything: discovery is a cycle, and each part of the book captures a different entry point into that cycle. Software engineering discovers requirements and architectures; data and models discover patterns and laws; knowledge systems discover connections across the literature; simulations discover consequences of hypotheses; and autonomous systems close the loop by running the entire cycle without pause. The book is sequenced so that each new entry point builds on the machinery of the ones before it. The next page describes who this journey is designed for.