This book sits at the intersection of three fields that rarely share a bookshelf: software engineering, machine learning, and scientific research methodology. It is written for people who work in at least one of those fields and need to reach into the others. Three reader profiles shaped every chapter.
The Software Engineer Building AI-Powered Research Tools
You ship code for a living. You have wired up LLM APIs, built retrieval pipelines, and you are increasingly asked to build systems that reason about code, literature, or scientific data. You want to understand how coding agents actually work, how to construct reliable agentic workflows, and how to connect AI tools to scientific databases and instruments.
Part II is your home base. Chapters 7 through 24 cover the full arc of AI-assisted software engineering: from vibe coding as a disciplined four-step loop of specification, steering, verification, and repair (Chapter 9) through context engineering at repository scale (Chapter 11), building MCP servers for scientific workflows (Chapter 12), multi-agent software teams (Chapter 17), and the trajectory toward autonomous software organizations (Chapter 24). Vibe coding in this book is not a shortcut; it is a discipline with formal specification, systematic verification, and structured repair.
The Discovery Workbench architecture in Chapter 6 gives you the platform design you need: components, data flows, artifact stores, agent loops, and evaluation harnesses. From there, the discovery system chapters in Parts IV and V supply the backend intelligence: retrieval-augmented discovery (Chapter 37), knowledge graph construction (Chapter 38), hypothesis generation (Chapter 39), and automated experiment design (Chapter 46). You will build each component, wire it into the Workbench, and evaluate it against established benchmarks.
The Graduate Student or Researcher
You read arXiv weekly. You need a foundation that lets the latest papers feel like an extension of what you already know rather than a disconnected set of tricks. You want the mathematical machinery: score functions, equivariant architectures, causal graphs, posterior predictive checks, all derived carefully from first principles.
Parts I, III, IV, V, and VI are the scientific core of the book. Part I (Chapters 1 through 6) lays the foundations: discovery as search through hypothesis space, knowledge representation with ontologies and property graphs, reasoning from deduction through abduction, and the three complementary discovery paradigms of data, models, and simulation. Part III (Chapters 25 through 35) builds the machine learning toolkit: scientific foundation models like ESM-3 and AlphaFold3 (Chapter 27), causal discovery and do-calculus (Chapter 31), Bayesian inference with calibrated uncertainty (Chapter 32), physics-informed neural networks and neural ODEs (Chapter 33), diffusion and flow-matching generative models with SE(3)-equivariant architectures (Chapter 34), and symbolic regression for equation discovery (Chapter 35).
Part IV (Chapters 36 through 41) covers the knowledge infrastructure: literature mining at scale, RAG for scientific corpora, hypothesis generation, research agents like ChemCrow and Atlas/ChemOS 2.0, and scientific claim validation. Part V (Chapters 42 through 47) adds the computational machinery: differentiable programming with JAX, world models for counterfactual reasoning, Bayesian optimization, and automated experiment design with active learning. Part VI (Chapters 48 through 52) deploys the full toolkit across biology, chemistry, physics, climate science, and social systems.
Graduate-depth sections throughout the book connect implementations to theory in both directions: from the mathematics to the code that realizes it, and from empirical behavior back to the principles that explain it. Appendix A maps three reading paths (Graduate Research, Software Practitioner, Domain Scientist) with chapter-by-chapter schedules.
The Domain Scientist Adding AI to a Research Workflow
You are a chemist, biologist, physicist, or climate scientist. You have heard claims about AI accelerating discovery, and you need to separate genuine capability from hype. You want to understand which AI tools are genuinely useful for your problems, how to evaluate their outputs rigorously, and how to build reproducible pipelines that produce defensible scientific claims.
Part VI covers your domain directly. Chapter 48 walks through AI for biology and medicine: drug discovery, genomics, single-cell analysis, and the AlphaFold3 case study anchored in the 2024 Nobel Prize in Chemistry. Chapter 49 covers chemistry and materials: molecular generation, retrosynthesis, reaction prediction, and the materials discovery pipeline with GNoME, MatterGen, and MACE-MP-0. Chapter 50 addresses physics and engineering: symbolic regression for physical laws, surrogate models, and neural PDE solvers. Chapter 51 treats climate and Earth science: climate emulators, extreme weather prediction, and the AI4Science agenda. Chapter 52 handles social and economic systems, with the unique methodological challenges of discovery in complex adaptive systems.
Parts III through V supply the methods behind the domain applications. The causal discovery chapter (31) explains when correlations in your data can support causal claims and when they cannot. The Bayesian uncertainty chapter (Chapter 32) shows how to quantify what your model does not know. The automated experiment design chapter (Chapter 46) shows how to choose the next experiment based on current knowledge and uncertainty, closing the loop between AI recommendation and laboratory execution.
The Discovery Workbench, introduced in Chapter 6 and extended through every subsequent chapter, gives you a practical architecture for a human-supervised AI research assistant: a system where you set the scientific goals, the AI proposes experiments and analyzes results, and you retain full control over which claims reach a manuscript.
What You Need Before Page One
Three prerequisites, all common in the target audience. First, Python proficiency: you should be comfortable writing functions, classes, and scripts, installing packages, and working in a development environment. The code throughout the book favors clarity over cleverness. Second, basic linear algebra: vectors, matrices, dot products, eigenvalues, and comfort with the idea that a transformation can be written as a matrix. Third, comfort reading mathematical notation: summations, integrals, probability distributions, and the conventions of mathematical writing. You do not need to prove theorems; you need to follow derivations and connect them to code.
What You Do Not Need
You do not need prior machine learning or deep learning experience. The book builds every model from components, starting from the mathematical definition and ending with working code and evaluation. You do not need prior scientific domain expertise: each domain chapter (Part VI) opens with the scientific context a newcomer needs, and graduate-depth sections are clearly marked for readers who want to go further. You do not need experience with LLM APIs or agent frameworks: Part II builds the entire agentic coding stack from first principles, starting with what an LLM call returns and ending with autonomous multi-agent systems.
What This Book Is Not
It is not a survey of the research literature; each chapter curates an annotated bibliography rather than citing everything published. It is not a cookbook of recipes for one library version; APIs appear throughout, but the goal is the understanding that survives the next major release. It is not a popular-science tour of AI breakthroughs; every claim is backed by code you can run and results you can reproduce. And it is not a replacement for domain expertise; the book gives you the AI tools and the methodology to deploy them, but the scientific judgment about what constitutes a meaningful discovery in your field remains yours.
If that matches your situation, continue to the next page for a guided preview of what the chapters look like inside, or jump to Appendix A to find the reading path designed for your background.