Introducing 'Accelerating Astrophysical Discovery'

Accelerating Astrophysical Discovery is a consortium effort with two connected goals.

The first goal is to design how future astrophysicists and cosmologists will do science in collaboration with AI. This means more than adding chat interfaces to existing tools. It means redesigning scientific workflows so humans, reasoning language models, simulators, data archives, instruments, and foundation models can jointly form hypotheses, retrieve evidence, run checks, update beliefs, and communicate results.

The second goal is to build the technical substrate that makes this possible: a joint-representation model that understands physical data. Such a model should learn across observations, simulations, synthetic observations, language, code, and instrumental context. It should support embeddings for retrieval and serendipity, and generative capabilities for conditional inference across linked views of the same physical system.

Discovery needs more serendipity

The main driver of discovery is serendipity. Modern astronomy now produces high-quality data faster than we can turn it into understanding. The bottleneck is no longer only data collection. It is knowing what to notice, what to compare, what to retrieve, and what surprising relationships deserve human attention.

A joint-representation model can make serendipity operational. If the embedding space is physically meaningful, it can surface rare candidates, unexpected analogues, counterexamples, and neighbouring evidence that a human may not know to ask for. The aim is not to replace scientific judgement, but to widen the field of attention.

Discovery also needs experimental discipline

The second main driver of discovery is rigorous control of experimental design. Scientific progress depends on precise protocols, careful control of priors and uncertainties, and a Bayesian framework for updating hypotheses as evidence accumulates.

AI systems must be designed around that discipline. They should not simply generate plausible explanations. They should expose assumptions, preserve uncertainty, keep track of negative results, and make it clear which data support which hypothesis. Generative capability is scientifically useful only when it remains connected to protocols, controls, forward models, and validation.

Papers are not the primary object

Publish or perish should become a thing of the past. Papers are subjective human summaries. They are essential for communication, but they are not the most fundamental scientific object.

In an AI-native scientific workflow, the central objects are hypotheses, evidence, protocols, and results. Papers become derivations and explanations built from those objects. Results should have durable identifiers. Hypotheses should be explicitly linked to the data, simulations, code, priors, assumptions, and validation steps that support or weaken them. Scientific protocols should reference those objects directly rather than relying only on narrative summaries.

Science was not built to share with AI

The current scientific framework was not designed to share the road with AI. A useful analogy is a city built only for pedestrians into which cars are suddenly introduced. If both are to coexist safely and productively, parts of the city need to be redesigned: roads, sidewalks, signals, rules, and shared expectations.

Scientific infrastructure faces a similar transition. Our papers, archives, code, provenance records, review systems, and collaboration norms were built for human-only workflows. AI systems can already read and generate language, but science requires more than language. It requires traceable relationships between data, hypotheses, simulations, instruments, uncertainty, and claims. We need scientific infrastructure that lets humans and machines reason together without blurring responsibility.

Physical data needs its own structure

Astrophysical data are heterogeneous and relational. A galaxy image, gravitational-wave strain segment, spectrum, catalogue row, simulation snapshot, or instrument response may describe a shared physical system, but these data do not form one natural sequence. Their relationships can depend on source, position, time, scale, measurement process, simulation history, and provenance.

This matters for AI because the structure used to learn from physical data must express those relationships. Language contributes composable concepts; physical data grounds them in evidence, measurement, simulation, uncertainty, and physical constraints. A useful joint representation must learn across both.

The role of the joint-representation model

The joint-representation model is the bridge between these ideas. It should learn when different data products refer to the same physical situation, when they are informative about one another, and when a generated completion is physically admissible rather than merely plausible.

We envision the relevancy graph as a large graph database in which nodes are physical data and links encode relevancy between them: the same source, sky region, time window, physical scale, instrument response, simulation-to-observation path, or provenance relationship. It is designed as training structure, not merely as a catalogue of associations.

In autoregressive language modelling, token order supplies a causal attention mask. Physical data have no universal left-to-right ordering, so self-supervised training needs a more general attentional field. When the system samples connected data from the graph, that connectivity becomes the attention mask: it determines which data can attend to one another while the model predicts, reconstructs, or aligns withheld physical information.

The long-term aim is an agentic discovery system in which reasoning LLMs guide retrieval, tool use, hypothesis generation, simulation, validation, and communication, while the joint-representation model provides grounded physical understanding. Humans remain responsible for scientific judgement, but they work with systems that can search, connect, and test evidence at a scale no individual can manage.

Key Problem Areas

The goals above are broad. We currently use four Key Problem Areas (KPAs) to make some of the scientific capabilities of a joint-representation model concrete. These KPAs are only a starting point: a working framework, not four separate projects or a closed list. The Leiden workshop and the consortium's continuing work will test whether they frame the right problems and may refine, combine, replace, or add KPAs.

These areas describe things that a joint-representation model could help scientists do. They may also provide training and evaluation targets for learning rich embeddings across observations, simulations, synthetic observations, language, code, and instrumental context. The aim is not to assign one loss function to each KPA, but to use the scientific problems to identify what information the representation must retain and what relationships the model must learn.

KPA 1: serendipity through embeddings

A physically meaningful embedding space could make serendipity operational. Relevant neighbours in that space could support rare-object discovery, anomaly triage, counterexample search, similarity retrieval, and context discovery that a scientist may not know to request.

Training targets could reward the model for aligning related views of the same physical system, retrieving withheld relevant data, and distinguishing superficially similar but physically different examples. Evaluation should ask whether the resulting embeddings surface rare, surprising, or scientifically meaningful candidates faster and more reliably than current workflows.

KPA 2: generative inference across instruments

Every instrument sees the universe through limits in resolution, spectral coverage, cadence, noise, calibration, sensitivity, and selection. A joint-representation model could learn across these limits to support deconfusion, gap filling, missing-modality prediction, cross-instrument reconstruction, and other forms of conditional inference without pretending that missing information is known.

Held-out modalities, controlled degradation, uncertainty calibration, forward-model checks, and comparison with classical baselines could serve as training or evaluation targets. They test whether the representation captures shared physical information rather than only producing plausible-looking completions.

KPA 3: observations to physical states

Multimodal observations could be used to constrain candidate physical states or simulation initial conditions. Model-generated candidates can be evolved through system and instrument simulators and compared back with the observations, making simulation part of the learning and validation loop.

Forward-model consistency provides a strong target for representation learning. The model should retain enough physical information to propose admissible states, expose uncertainty, and reveal through residuals when conditioning information or relevant physics is missing.

KPA 4: human-machine discovery

The fourth area considers the full workflow: how scientists, reasoning LLMs, tools, simulators, archives, and the joint-representation model work together to form, test, revise, and communicate hypotheses. The model supplies grounded embeddings and generative physical inference; reasoning systems orchestrate retrieval and tools; humans remain responsible for scientific judgement.

Here, the targets extend beyond a model benchmark. We should test whether the combined workflow improves scientific search, triage, hypothesis generation, validation, or communication while preserving uncertainty, negative results, provenance, and human decision points.