Unraveling the Cosmos: Key Moments from the Annual Conference

The first annual conference for the Center for Decoding the Universe, on June 5-6 (agenda), convened an impressive group of exceptional researchers. Check out the full recap, including on-demand videos and event photos―scroll down to view the photo album.

Session 1: Time-Domain Data & Anomaly Discovery

The opening session on Time-Domain Data & Anomaly Discovery brought together perspectives from astrophysics, statistics, high-energy physics, and machine learning.

Ashley Villar discussed the upcoming revolution in time-domain astrophysics, with the Rubin Observatory expected to drive the supernova discovery rate from the current ~10,000 per year to one million per year. She grouped the science challenges by anomaly type and demonstrated how each demands a different feature space. Villar highlighted “one‑in‑a‑million” opportunities such as kilonova‑like collapsars and Population III explosions, while cautioning that rare events can hide in familiar brightness–timescale coordinates. She concluded with a call for rapid, data-driven inference pipelines capable of triaging millions of nightly alerts without missing the exotic discoveries.

The next speaker, Ben Nachman, reframed anomaly detection through a dual lens of latency (online vs. offline) and supervision (unsupervised to weakly supervised). Using techniques from particle physics, he contrasted density-based autoencoders, which are powerful but coordinate-dependent, with likelihood-ratio methods that remain invariant under nonlinear transformations. Nachman mentioned the case of the Gaia stellar streams, where a weakly supervised classifier automatically recovered the GD‑1 stream and its fine sub‑structure, all without explicit astrophysical priors.

Josh Bloom then asked a meta‑question: What do we want from anomaly detection? After showcasing a multimodal foundation model that represents light curves, spectra, and metadata into an interpretable embedding, Bloom discussed how many discoveries may be due to poor spectral reductions or photometric artifacts. He argued for pushing anomaly detection upstream, directly onto raw images and telemetry, to flag systematics before they appear as new physics. Bloom sketched a future in which automated Occam filters, key physics checks, and agentic LLMs do the first pass, leaving humans to interrogate only those outliers that survive the filters.

For the short talks, Maja Jablonska introduced SPICE. This JAX-based, fully differentiable forward model tessellates stellar photospheres and can render the spectra of spotted, pulsating, or eclipsing stars. Peter Melchior then presented a rest‑frame–aware auto‑encoder for galaxy spectra that, in just six latent parameters, flags broad‑absorption‑line quasars and star‑burst outliers while remaining invariant to redshift. Next, Charlotte Ward introduced Scarlet2, a JAX rewrite of the Rubin de‑blender that simultaneously fits Rubin, Euclid, Roman, and LS4 images. Linnea Wolniewicz then discussed a Gaussian-process pipeline to model complex stellar variability, enabling searches that capture hour-long dips and week-long exocomet occultations in Kepler light curves. Finally, Christine Ye described ResearchBench, an open benchmark that converts 22 peer-reviewed astrophysics papers into 121 autonomous reproduction tasks, with the frontier models scoring below 20%, revealing subtle but systematic reasoning gaps that future agents must address.

During the long discussion, participants debated the balance between domain knowledge and flexible ML. Several groups championed a version-controlled archive of recurring detector or reduction artifacts, against which every new alert could be cosine-matched. Resource efficiency surfaced as a priority, with one table proposing a three‑tier triage: physics‑encoded heuristics first, lightweight ML next, and only then deep learning. The room agreed that models should have saliency maps so humans can see why a detection occurred. Participants from NVIDIA and Stanford Medicine noted that low-frame-rate MRI sequences face identical rare-event problems and formed a cross-disciplinary working group. Finally, moderator Phil Marshall polled the room on who would find Rubin’s first novel discovery: humans received thirty‑five votes, ML twenty, and hybrid systems five.

The one‑minute poster session ranged from U‑map clustering of Chandra X‑ray transients to density‑ablation clustering for direct dark‑matter events.

Session 2: Cosmology & Modern Inference Frameworks

The second session, Cosmology & Modern Inference Frameworks, shifted the focus from discovery to precise parameter estimation when data volumes require more than classical statistics. Ben Wandelt (in a pre‑recorded keynote) surveyed the state of “implicit inference’’: likelihood‑free, simulator‑based techniques able to recover full cosmological posteriors even when the forward model includes expensive N‑body or hydrodynamic solvers. He discussed three open bottlenecks, namely scalability, model fidelity, and mis‑specification, and showed how transformer‑based halo emulators (Gotham) and diffusion‑style field reconstructions can close each gap. Wandelt closed with a challenge: prove that tomorrow’s Rubin‑scale analyses can propagate all sources of uncertainty, from sub‑grid baryonic recipes to adversarial domain shifts in simulation code.

Laurence Perreault Levasseur focused on a concrete use case: joint lens‑mass + source‑reconstruction for the hundreds of thousands of galaxy‑scale strong lenses Rubin will deliver. She argued that classical pixel‑based inversions struggle when using 10,000+ nuisance parameters. Her group replaces the hand‑tuned Gaussian priors with score‑based generative models that learn the manifold of high‑redshift galaxies. In simulation, the approach removes the size‑bias that a zero‑centred Gaussian prior imprints on lens‑mass estimates and remains stable when the noise model itself is learned from JWST dark frames via diffusion networks.

The session’s third talk came from Vasilis Syrgkanis, who bridged theory and practice by discussing automatic debiasing. He mentioned that semi‑parametric estimators often inherit the regularisation bias of the machine‑learning models that supply their nuisance functions. Syrgkanis demonstrated the recipe on causal‑effect estimation and outlined open problems at the cosmology–statistics interface, notably high‑dimensional hierarchical models.

For the short talks, Sydney Erickson focused on a neural posterior estimator that delivers per‑lens mass posteriors from HST‑like images, which are then fed into a population‑level cosmology fit that simultaneously learns the lens‑galaxy mass function. Tiffany Fan presented a context‑aware graph neural surrogate for particle‑accelerator beams. Sophia Lu introduced ABI (Adaptive Bayesian Inference), a likelihood‑free rejection‑ABC variant that refines posterior quantiles via a Wasserstein distance. Guillem Megias Homar discussed early Rubin commissioning data, and Shuo Xin’s talk focused on neural rendering that employs adaptive mesh refinement.

The long discussion centered around two key issues. First, if simulator‑based pipelines no longer fit individual images “down to the noise,” how do we certify that the aggregated cosmological posteriors are still trustworthy? Suggestions ranged from blinded end‑to‑end challenges to per‑statistic saliency checks that reveal which scales drive the inference. Second, everyone discussed the model and prior mis‑specification. Consensus emerged that the community needs (i) flexible, data‑driven priors that can be updated hierarchically as new samples roll in, and (ii) robustness tests that quantify how far wrong the forward model can be before the posteriors break. The session closed with a call to develop public mis‑specification test‑beds, analogous to ImageNet adversarial suites, so methods can be stress‑tested before Rubin data releases.

One‑minute poster talks discussed diffusion‑based deconvolution for SN Ia host spectroscopy, negative‑weight “neuro‑refinement’’ for collider unfolding, and U‑MAP clustering of red‑shift–space galaxy mocks.

Session 3: Galaxies and Foundation Models

Marc Huertas‑Company opened by discussing the traditional, survey‑driven approach to decode galaxy evolution: derive physically motivated summary statistics, search for scaling relations, and tune cosmological simulations. He showed how neural networks already replace many handcrafted estimators (stellar mass, SFR, metallicity) but warned that familiar correlations still allow vastly different sub‑grid physics. He argued that the next step is multi‑modal, multi‑scale AI: joint embeddings of optical, radio, X‑ray, and simulation data that map models and observations into a higher‑dimensional space, which is the only way to break degeneracies in gas accretion, feedback, and quenching.

François Lanusse discussed the methodological change from single‑task supervised CNNs to reusable foundation models. Self-supervised contrastive learning on hundreds of millions of images yields embeddings that perform as well as or better than existing models for lens finding, morphology, redshift, or stellar‑mass estimation. Extending CLIP‑style training to spectra‑plus‑images (AstroCLIP) creates a shared space where a nearest neighbour in embedding space transfers physical parameters with state‑of‑the‑art accuracy. He showcased a generative transformer that tokenizes every modality (HST, JWST, Euclid, SDSS) and can impute high‑resolution spectra or multi‑band images from sparse inputs, which is a step toward “an astronomical GPT” that both simulates and interprets the sky.

Drawing on 15 years in computer vision, David Fouhey put the spotlight on self‑supervised depth and segmentation breakthroughs, yet cautioned that they required vast, carefully curated benchmarks. Like autonomous‑driving edge‑cases, spiral disks at z ≈ 10 or solar mega‑flares may appear in the long tail where embeddings and uncertainty quantification may not be enough. Fouhey urged the community to invest in rigorous downstream tests to keep plausible pictures from substituting for real understanding.

A lively discussion followed on how to insert physics into learned representations, whether we will soon work only in embedding space, and how to detect novel phenomena that foundation models have never shown.

The short talk session started with Philipp Frank, who introduced Geometric Variational Inference, which is already mapping Milky‑Way dust and black‑hole shadows. Sonia Kim demonstrated an inverse‑problem solver that is competitive with state‑of‑the‑art MAP and posterior samplers. Benjamin Remy showed how iterative diffusion priors sharpen source‑separated HST images. Rahul Mysore Venkatesh then discussed motion‑completion transformer predicts object‑wise causal masks from raw video, hinting at similar embeddings for time‑domain surveys. Finally, Sebastian Wagner-Carena presented a multi‑view diffusion framework that learns galaxy‑and‑background priors without labels, achieving clean component separation across resolutions.

Poster spotlights ranged from physics‑guided diffusion guidance to vision‑transformer photometric redshifts and Roman–LSST deblending methods.

The takeaway is that foundation models are rapidly becoming an integral part of survey astrophysics, but extracting physics from their embeddings will demand equal parts creativity, rigor, and domain knowledge.