2026/06/14 by Ran Liu · 1 voice
paper · doi:10.5281/zenodo.20691098
openalex publication_date 2026/06/14 · openalex created_date 2026/06/15 · openalex updated_date 2026/07/01
Astrophysics is entering a stage in which discovery increasingly depends on large-scale surveys, long-lived archives, real-time alert streams, multi-wavelength and multi-messenger evidence, high-dimensional simulations, and inference pipelines that exceed what can be organized through manual inspection alone. AI is already being used across this landscape: to process observations, detect and rank candidates, search archives, learn transferable representations, accelerate inference, emulate simulations, and support theory-facing analysis. Yet the conceptual organization of AI-for-astrophysics work remains fragmented. Projects are often described by subfield, instrument, dataset, model family, or task, making it difficult to compare what kind of astrophysical contribution is being made and how deeply AI is actually integrated into the discovery process. This fragmentation creates a strategic problem for researchers and system designers. A classifier embedded in a survey pipeline, a cross-modal representation model, an AI-assisted anomaly search, a simulation emulator, and an agentic observatory prototype may all be called “AI for astrophysics,” but they occupy very different positions in the scientific system. They differ in the domain aspect they intervene in, the level of AI integration they require, the core AI functions they use, the scientific value they pursue, and the maturity of the regime they belong to. Without a shared map, AI adoption can become model-driven rather than science-driven, and promising work can remain difficult to compare, extend, or evaluate. The AI for Astrophysics Domain Compass addresses this problem by providing a domain-grounded framework for locating, interpreting, designing, and comparing AI integration in astrophysical discovery. It treats astrophysics as an observational, inferential, and explanatory discovery system: cosmic signals become validated evidence; evidence is organized into objects, events, populations, and systems; and these are explained through physical mechanisms, evolutionary pathways, and fundamental constraints. The Compass combines this domain structure with AI integration depth, AI core functions, a regime map, and a value compass. It is designed to support both research mapping and system design without reducing the field to a list of AI tools, model architectures, or isolated use cases. What the Compass enables Domain-grounded positioning.The Compass allows an AI-for-astrophysics project to be located according to the primary astrophysical role it serves: evidence formation, phenomenon organization, or physical explanation. This helps distinguish projects that may use similar AI methods but intervene in different parts of the scientific process. Integration-depth clarity.The framework separates shallow interaction with existing AI from workflow embedding, domain adaptation, and architecture-level reinvention. This makes it possible to ask what technical expertise, data infrastructure, validation responsibility, and system redesign are required for a project to move from one tier to another. Mechanism-level description.By using the AI Core Functions layer, the Compass describes AI systems according to what capabilities they exercise in the astrophysical context: representation, inference, optimization and control, simulation and emulation, generation, and orchestration. This provides a model-agnostic vocabulary for mechanism design and comparison. Regime-level navigation.The AI for Astrophysics Regime Map identifies twelve AI integration regimes across the three astrophysical domain aspects and four integration tiers. Each cell names a broad system family, such as Intelligent Observation-to-Evidence Pipelines, Transferable Discovery Models, or Physical Model-Building Platforms. The map helps users locate current work, identify underdeveloped regimes, and plan movement toward deeper or more ambitious forms of integration. Value-sensitive evaluation.The Value Compass grounds AI integration in astrophysics-intrinsic values: Cosmic Significance, Scientific Credibility, and Cumulative Discovery Capacity. These values help researchers and designers clarify whether a project primarily advances consequential cosmic understanding, strengthens the credibility of scientific outputs, or builds reusable capacity for the astrophysics community. Components of the Compass and their roles 1. AI–Astrophysics Space The AI–Astrophysics Space provides the high-level coordinate system for the framework. A project can be positioned by asking three questions: which astrophysical domain aspect it primarily affects, how deeply AI is integrated, and which value direction it most strongly serves. This makes the framework useful for locating individual papers, comparing research streams, planning system development, and identifying strategic gaps. The space is not intended to replace existing astrophysical subfields. Cosmology, stellar astrophysics, galaxy evolution, exoplanet science, compact objects, time-domain astronomy, and multi-messenger astrophysics can all be mapped into the same space. The framework organizes contributions by epistemic role rather than by object class or observational modality. 2. Astrophysical domain aspects The three domain aspects define the stable astrophysical structure of the Compass. They represent the major roles through which astrophysical knowledge is produced: acquiring and validating evidence, organizing cosmic phenomena, and constructing physical explanation. Each aspect contains finer-grained subcategories that support more precise positioning, but the main purpose of the aspect layer is not to create a complete taxonomy of astrophysics. It provides a stable skeleton for identifying the primary contribution of a method, system, paper, or research program. For research mapping, the domain aspect layer can be used to identify whether a paper primarily improves evidence production, discovers or organizes phenomena, or advances physical interpretation. Secondary aspects can also be recorded when a work bridges multiple parts of the discovery system. 3. Depth of AI Integration The integration-depth axis adapts the Four-Tier Framework for Human–AI Collaboration (DOI: 10.5281/zenodo.17701769). In this Compass, the four tiers describe increasing design responsibility in the use of AI for astrophysics. At Tier 1, existing AI supports user-initiated work. At Tier 2, existing AI becomes a bounded component in an established workflow. At Tier 3, AI is adapted with domain data, constraints, or objectives to provide reusable field-specific capability. At Tier 4, new AI architectures and domain processes are designed together to meet requirements beyond existing systems or adaptation. This axis helps users distinguish interactive AI use from workflow embedding, domain specialization, and architecture-level reinvention. It is especially useful for planning because movement across tiers usually requires different levels of AI expertise, validation responsibility, data infrastructure, and system design commitment. 4. AI Core Functions The mechanism layer draws on the AI Core Function Ontology (DOI: 10.5281/zenodo.17664037). It explains how AI acts within the astrophysical discovery system by identifying the functional roles that AI capabilities play in evidence formation, phenomenon organization, and physical explanation. This layer keeps the Compass model-agnostic. The same model family may serve a representation function in one regime, an inference function in another, or a generative function in a third. Conversely, a single astrophysical system may combine several functions. For research mapping, the function layer helps describe what an AI contribution actually does. For system design, it helps identify which capabilities are needed, which can be supplied by existing AI, and which require domain adaptation or new architecture. 5. Value Compass The Value Compass defines the domain-intrinsic value directions that guide AI integration in astrophysics. These values do not come from AI itself. AI changes how they are pursued, but the values remain grounded in the scientific aims and shared sicentific practices of astrophysics. Cosmic Significance evaluates whether a contribution expands, deepens, or reframes scientific understanding of the universe. Scientific Credibility evaluates whether scientific outputs are well supported, robust under relevant uncertainties, and traceable for independent scrutiny. Cumulative Discovery Capacity evaluates whether a contribution builds shared resources, tools, standards, or infrastructure for cumulative use by the astrophysics community. This layer is important because different AI systems can occupy the same regime while pursuing different values. A discovery engine optimized for rare-object search may prioritize Cosmic Significance. A pipeline designed around uncertainty, provenance, and calibration may prioritize Scientific Credibility. A shared model, archive, benchmark, or workflow infrastructure may prioritize Cumulative Discovery Capacity. Making the value emphasis explicit allows trade-offs to be discussed before technical design choices become locked in. 6. AI for Astrophysics Regime Map The Regime Map is the 4×3 grid formed by Integration Tier × Astrophysical Domain Aspect. Each cell names a broad AI-for-astrophysics regime family. These regime families are deliberately model-agnostic. They are meant to cover classes of systems and research directions rather than individual methods or specific architectures. The maturity shading in the map should be read as a qualitative indication of practical development and institutionalization, not as a normative ranking. A mature regime can still contain unresolved credibility or governance problems. A low-maturity regime can be scientifically important but technically premature. The map supports field-level orientation by showing where current