Deep Learning for Astrophysics: An Open Textbook from the NASA Cosmic Origins AI/ML Science and Technology Interest Group

Yuan-Sen Ting, Digvijay Wadekar, Phill Cargile, Carol Cuesta-Lazaro, André Curtis-Trudel, Gregory Green, Ryan McClelland, Daniel Muthukrishna, Tri Nguyen, Helen Qu, Tomasz Rozanski, Anna Scaife, Jesse Thaler, Licia Verde, Francisco Villaescusa-Navarro, John F. Wu, Duo Xu, Siyu Yao, Alex Gagliano, Siddharth Mishra-Sharma, Andrew K. Saydjari, Georgios Valogiannis, Peter Kurczynski, Swara Ravindranath · 2026

Read Deep Learning for Astrophysics: An Open Textbook from the NASA Cosmic Origins AI/ML Science and Technology Interest Group by Yuan-Sen Ting, Digvijay Wadekar, Phill Cargile, Carol Cuesta-Lazaro, André Curtis-Trudel, Gregory Green, Ryan McClelland, Daniel Muthukrishna, Tri Nguyen, Helen Qu, Tomasz Rozanski, Anna Scaife, Jesse Thaler, Licia Verde, Francisco Villaescusa-Navarro, John F. Wu, Duo Xu, Siyu Yao, Alex Gagliano, Siddharth Mishra-Sharma, Andrew K. Saydjari, Georgios Valogiannis, Peter Kurczynski, Swara Ravindranath (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-30855.txt.

Introduction

Yuan-Sen Ting and colleagues present an open textbook, Deep Learning for Astrophysics, grown from a NASA interest group's lecture series. Community assessments name education, not access to tools, as the main barrier to astronomers adopting machine learning. The NASA Cosmic Origins AI/ML Science and Technology Interest Group formed to close that gap with short, field-specific teaching, and this book collects its lectures for free online.

The textbook holds 23 chapters in six parts, drawn from 17 lecturers. It runs from computing foundations and network architectures through generative modelling, simulation-based inference and reinforcement learning to language-model agents, closing with the practice of AI-assisted science. Many chapters carry executable notebooks. The group also sketches its next year, centred on agentic research and NASA's ASTRA mission-concept initiative.

Catalog record

CREATORYuan-Sen Ting, Digvijay Wadekar, Phill Cargile, Carol Cuesta-Lazaro, André Curtis-Trudel, Gregory Green, Ryan McClelland, Daniel Muthukrishna, Tri Nguyen, Helen Qu, Tomasz Rozanski, Anna Scaife, Jesse Thaler, Licia Verde, Francisco Villaescusa-Navarro, John F. Wu, Duo Xu, Siyu Yao, Alex Gagliano, Siddharth Mishra-Sharma, Andrew K. Saydjari, Georgios Valogiannis, Peter Kurczynski, Swara Ravindranath
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE11 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.30855

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Deep Learning for Astrophysics: An Open Textbook from the NASA Cosmic Origins AI/ML Science and Technology Interest Group Yuan-Sen Ting, Digvijay Wadekar, Phill Cargile, Carol Cuesta-Lazaro, André Curtis-Trudel, Gregory Green, Ryan McClelland, Daniel Muthukrishna, Tri Nguyen, Helen Qu, Tomasz Rozanski, Anna Scaife, Jesse Thaler, Licia Verde, Francisco Villaescusa-Navarro, John F. Wu, Duo Xu, Siyu Yao, Alex Gagliano, Siddharth Mishra-Sharma, Andrew K. Saydjari, Georgios Valogiannis, Peter Kurczynski, Swara Ravindranath Abstract The astronomical community’s ability to use modern machine learning shapes the science return of upcoming facilities. Recent community assessments single out education as the principal barrier to adoption, because what limits uptake is the uneven understanding of these methods rather than their availability. The NASA Cosmic Origins Artificial Intelligence and Machine Learning Science and Technology Interest Group (AI/ML STIG) was formed to address this gap through short, domain-specific tutorials and community discussion. We present Deep Learning for Astrophysics , a freely available textbook at https://deeplearning4astro.com , curated from the group’s lecture series. It collects 23 chapters across six parts from 17 lecturers, running from computational foundations and deep-learning architectures through generative modeling, simulation-based inference, and reinforcement learning to large-language-model agents, and closing with the practice of AI-laden science. Many chapters include executable notebooks. We also outline the group’s plan for the coming year, centered on agentic research and on NASA’s ASTRA mission-concept initiative. I The gap is adoption, not availability The scientific case for machine learning in astronomy is established. Data from the Vera C. Rubin Observatory and the Nancy Grace Roman Space Telescope will strain traditional analysis, and learned methods already aid classification, inference, and instrument operation. Uptake remains the limiting problem. The NSF AI+MPS community paper ( 1 ) sets out strategic priorities for the mathematical and physical sciences, one of them to “foster education and workforce development in AI.” It names an undertrained community, distrust of opaque methods, and resources written by and for computer scientists as barriers, and proposes modular, stackable training in response. A review of deep learning in astrophysics ( 3 ) makes a complementary point. The value of these methods rests on understanding why they work, on architectural choices that encode physical assumptions and on testability, rather than on predictive accuracy alone, and progress depends on teaching deep learning as an extension of statistical practice rather than a replacement for it. Understanding is the prerequisite for trust, and trust for adoption. This matters because AI can produce an appearance of understanding without the substance of it ( 2 ) . The community paper also flags a feedback loop: undertraining can lead to work with weak methodological grounding, which hardens the skepticism it provokes. Raising literacy can improve both the uptake and the quality of AI in the field. Literacy matters, too, for those who never train a model. Astronomers referee papers, sit on panels, and edit journals, and as AI-generated work enters the literature, understanding the methods is what lets them separate scientific insight from plausible output. The field needs researchers able to judge the work, not only to produce it. Both assessments point to the same response: domain scientists, not only computer scientists, should build AI training tailored to their field, with domain-relevant examples and criteria for when a learned method is the right tool. II The interest group NASA’s Cosmic Origins program supports the community studying how galaxies, stars, and cosmic structure form and evolve, with missions and concepts spanning the Hubble Space Telescope and the future Habitable Worlds Observatory. The AI/ML STIG is a Science and Technology Interest Group of its Program Analysis Group (COPAG), open to the national and international community. 1 1 1 NASA AI/ML STIG home page at https://science.nasa.gov/astrophysics/programs/cosmic-origins/community/artificial-intelligence-machine-learning-science-technology-interest-group-ai-ml-stig/ . Its premise is that training the community can increase the science return of NASA’s missions, and that the training must be domain-specific and time-efficient. The design responds to obstacles named in the community paper and the group’s founding proposal. These include resistance to AI among established researchers, the difficulty of assessing interdisciplinary skill, faculty time constraints that make full courses impractical, fragmented resources that lack astronomical context, and a pace of change that outruns standard curricula, with LLM agents the current example. The response is a training sequence assembled from stackable short modules, each covering a self-contained competency. The aim is not to turn astronomers into machine-learning engineers but to help researchers connect astronomical questions with machine-learning methods. To that end, the group runs short tutorials on individual methods, with time for community questions and concerns about reliability. A leadership council of seven organizes the activities. 2 2 2 The author list is ordered as co-chairs, then lecturers, then the leadership council, then the Cosmic Origins program officers, with each group alphabetical by surname. Yuan-Sen Ting is both a co-chair and a lecturer, and Carol Cuesta-Lazaro is both a lecturer and a member of the leadership council. As a COPAG STIG, the group also reports community needs and technology gaps through COPAG channels. III The textbook Figure 1: The web version of Deep Learning for Astrophysics . The textbook turns the AI/ML STIG lecture series into modular, domain-specific chapters that give astronomers practical entry points into modern machine-learning methods. Over the past academic year the group ran an online lecture series. 3 3 3 Lecture series, recordings, and source materials at https://tingyuansen.github.io/NASA_AI_ML_STIG/ . We have curated these lectures into a single textbook, Deep Learning for Astrophysics , available online. 4 4 4 https://deeplearning4astro.com . The lectures, notebooks, and slides were edited into a consistent chapter format and reviewed by the contributing lecturers. The executable notebooks and their outputs are preserved, so readers can run and modify the methods rather than only read about them (Figure 1 ). The exercises use real astronomical data, such as galaxy images, radio-galaxy morphologies, transient light curves, and stellar streams, rather than toy examples. Each notebook chapter pairs the minimum theory a method needs with a worked astronomical example, so that its uses and its failure modes are learned together. Of the 23 chapters, seventeen are executable notebooks and six are concise readings, grouped into six parts. The intended reader is a graduate student or researcher who wants working knowledge rather than a survey. Because the parts are self-contained, the book can be read end to end or used as standalone modules in a course. The broader lecture series also included sessions on the AI funding landscape and on AI-enabled mission concepts; the textbook focuses on the technical and interpretive material most suited to chapter format. Part I. Foundations • The Future of AI and the Mathematical and Physical Sciences (Jesse Thaler) • PyTorch and Autodifferentiation (Phill Cargile) • JAX (Phill Cargile) • Inductive Biases (John F. Wu) Part II. Deep Learning Architectures • Convolutional Neural Networks (John F. Wu) • Recurrent Neural Networks (Daniel Muthukrishna) • Transformers (Helen Qu) • Graph Neural Networks (Tri Nguyen) • Equivariant Neural Networks (Anna Scaife) Part III. Generative Models & Inference • Normalizing Flows (Gregory Green) • Diffusion Models (Duo Xu) • Flow Matching (Tomasz Rozanski) • Simulation-Based Inference (Tomasz Rozanski) Part IV. Reinforcement Learning • Reinforcement Learning (Carol Cuesta-Lazaro) Part V. Large Language Models & Agents • LLM API Basics (Yuan-Sen Ting) • RAG and

Read the complete text