Agenda

As global carbon markets rapidly expand to combat climate change, the integrity of carbon offsets projects, ranging from reforestation to wetland restoration, remains a critical bottleneck. The viability of these markets heavily relies on robust Monitoring, Reporting, and Verification (MRV) to ensure additionality and accurately quantify carbon sequestration. However, traditional MRV relies on expensive, unscalable, and infrequent on-site measurements.

In this talk, we will outline the methodology of using remote sensing with image classification to monitor ecological projects at scale. This approach not only allows us to verify that projects are progressing as claimed, but also enables the calculation of dynamic carbon discounts based on a project's probabilistic risk of failure.

Crucially, we will address the data-labeling bottleneck inherent in this work. Because obtaining ground-truth measurements is cost-prohibitive, training traditional models is difficult. To overcome this, we explore the emerging frontier of Geospatial Foundation Models. We will discuss how these powerful models enable few-shot learning, allowing us to extract high-confidence insights with drastically fewer labeled measurements.

Stellar collisions occur frequently in dense environments, and play a key role in producing exotic phenomena from blue stragglers in globular clusters to high-energy transients in galactic nuclei. Successive collisions of massive stars could lead to the formation of massive black holes, serving as seeds for supermassive black hole in the early universe. While analytic fitting formulae exist for predicting collision outcomes, they do not generalize across different energy scales or stellar evolutionary phases. Smoothed particle hydrodynamics (SPH) simulations are often used to compute the outcomes of stellar collisions, but, even at low resolution, their computational cost makes running on-the-fly calculations during an N-body simulation challenging. We present a new grid of 27,720 SPH calculations of main-sequence star collisions, spanning a wide range of masses, ages, relative velocities, and impact parameters. Using this grid, we train machine learning models to predict collision outcomes (merger vs disruption, or flyby) and final remnant masses. We compare the performance of nearest neighbors, support vector machines, and neural networks, achieving classification balanced accuracy of 98.4%, and regression relative errors as low as 0.11% and 0.15% for the final stars 1 and 2, respectively. We make our trained models publicly available as part of the package collAIder, enabling rapid predictions of stellar collision outcomes in N-body models of dense star cluster dynamics.

The advent of broadband, high-spectral resolution receivers and correlators on telescopes like ALMA, NOEMA, and the SMA provides detailed maps of chemically rich regions which contain a full spectral line survey in each of many thousands of pixels. Next-generation facilities and upgrades like ALMA WSU, the ngVLA, and the SKA will greatly magnify these datasets. Yet, the analysis of even a single spectrum's chemical content is a process that can still take hours or days, even with some degree of automation. AstroAMASE is an automated line identification and prediction tool built on the framework originally built for laboratory analysis (AMASE: Automated Mixture Analysis via Structural Evaluation). I'll present an overview of this tool which can rapidly (minutes) and accurately (~98% validated assignments) determine linewidths, vlsr values, and assign most known chemical signatures including reasonable guesses at temperatures and column densities. Further, I'll discuss its ability to suggest other likely molecules to be present in the source, including those not in databases or with prior laboratory measurements. Finally, I'll touch on future directions connecting these efforts to ALMA data products through a Cosmic AI Seed Grant.

We present a self-supervised physics-informed neural network (PINN) framework for stellar structure modeling, motivated by the computational cost of traditional solvers in large-scale stellar population synthesis. This mesh-free approach learns solutions to the coupled stellar structure equations directly from physical constraints, without relying on labeled data. The network maps space-time coordinates of the star to key stellar quantities (enclosed mass, density, temperature, and luminosity profiles). The current implementation focuses on spatial solutions, while extensions to time-dependent evolution show that the PINN captures global trends (such as Hertzsprung-Russell diagram features), but remains limited for fine structural detail. To accurately recover stellar interiors, we incorporate state-of-the-art PINN enhancements, including hard boundary constraints, Fourier feature embeddings to resolve high spatial frequencies, a SIREN architecture, stochastic gradient evaluation, and active learning. Validation shows strong agreement with established finite-difference models across stellar masses, highlighting the potential for scalable and physically consistent stellar modeling.

Machine learning is rapidly transforming how we explore high-dimensional astronomical data, but how do we know when it is discovering new physics and when it is simply learning biases in the data or selection effects? In this talk, I present results from applying unsupervised learning to more than 200,000 galaxy spectra from the Sloan Digital Sky Survey (SDSS), using a Probabilistic Auto-Encoder (PAE) combined with UMAP to construct low-dimensional representations of spectral data. These representations reveal striking structure, including trends that appear to trace galaxy quenching pathways. However, they also reveal trends inherited from the survey design that are unrelated to intrinsic properties. At the same time, the model highlights multiple classes of outliers, spanning rare astrophysical objects as well as observational and data-processing artifacts. This duality between meaningful discovery and misleading signal marks a central challenge for AI in astronomy. Interpreting these results required a dedicated, iterative effort combining domain expertise, visualization, and detailed follow-up analyses. I will discuss what this process reveals about the strengths and failure modes of unsupervised ML. Some of these lessons can inform efforts to scale to much larger datasets from the Dark Energy Spectroscopic Instrument (DESI), containing >50 millions galaxies, as well as efforts to build next-generation astronomical foundation models.

We present a mathematically rigorous vector representation for galaxy morphological structure derived from multi-band astronomical imaging. Each galaxy is mapped to a compact, fixed-dimensional vector by partitioning its polar-coordinate field into radial and angular bins, computing pairwise distances between radial profiles across all angular shifts, and aggregating through an exponential kernel. The resulting signature is provably invariant under continuous spatial rotation, ensuring that physically identical galaxies receive identical representations regardless of their orientation. Furthermore, the mapping is invertible up to a precisely characterized, small finite set of equivalence-class transformations, providing near-lossless encoding of morphological information. Applied to Galaxy Zoo 2 survey data across $g$, $r$, and $i$ photometric bands, the descriptor enables efficient content-based galaxy retrieval: given a query galaxy, the nearest neighbors in signature space consistently recover visually and morphologically similar objects, effectively functioning as a search engine over large galaxy catalogs.  Using modern vector databases, these queries are extremely efficient (milliseconds), even over 100,000+ galaxies.  Moreover, because of the vector representation, this should in the future enable many other AI applications including clustering, classification, and anomaly detection -- as well as using richer spectral information. 

For large spectroscopic surveys like the Dark Energy Spectroscopic Instrument (DESI), accurate catalog redshifts are essential. However, standard methods produce unflagged failures. Identifying these errors without exhaustive manual inspection requires an independent estimator with calibrated uncertainty, p(z|spectrum).

Developing this estimator and modeling the underlying spectral density, p(spectrum), requires mapping high-dimensional spectra into a low-dimensional space. To extract physical meaning from these densities, the mapping must preserve a meaningful notion of distance between spectra. Standard autoencoders fail to satisfy this metric-preserving constraint.

We detail a two-stage approach built on metric learning to solve this. First, we construct an isometric latent space that directly enables highly accurate nonparametric inference, supporting k-NN regression and classification, similarity searches, and photometric redshift estimation. 

Second, we fit a Neural Spline Flow to this space to generate per-spectrum redshift distributions and anomaly scores. By flagging pipeline predictions that are low-probability when evaluated against the learned distribution, the model catches 60% of catastrophic failures while falsely flagging only 2% of good spectra. Because this evaluation is restricted to the pipeline's most confident estimates, this represents a lower bound on in-pipeline performance. Ultimately, this reduces the manual visual inspection workload by upward of 98%.

Tokenization is central to adapting scientific data for transformer-based foundation models, yet its impact on learned representations remains poorly understood. In this work, we present a systematic comparison of four tokenization strategies, Affine, AIM, JetFormer, and VQ-VAE, within a unified transformer framework for astronomical imaging. Using 640,000 galaxy images from the DESI Legacy Survey and a shared AstroPT backbone, we evaluate each method on reconstruction fidelity, physical property prediction, and the relationship between the two.

Our results reveal consistent trade-offs across approaches. The flow-based JetFormer achieves substantially higher reconstruction quality, recovering fine spatial structure and low-surface-brightness background features. The discrete VQ-VAE representations yield the strongest probe performance for galaxy physical properties, while the patch-based Affine and AIM tokenizers best preserve localized morphological information. Notably, reconstruction quality and representation quality are decoupled; no single method is universally optimal.

By grounding our evaluation in independently measured physical quantities, we demonstrate how scientific data, where known physics provides objective ground truth, can provide rigorous, interpretable benchmarks for foundation models.

We demonstrate applications of large language model (LLM) agentic systems to two challenges in computational astrophysics. First, we show that such systems can construct efficient surrogate models for expensive astrophysical simulations. Notably, these agents display an emergent capability to develop symbolic surrogates without any explicit symbolic regression design, enabling the discovery of an analytic surrogate for eccentric binary black hole mergers that is more than an order of magnitude cheaper than previously known solutions. Alongside this, we introduce an explicitly verifiable benchmark for surrogate model evaluation, providing a rigorous framework to measure and guide progress in this domain.

Second, we showcase the use of LLM agentic systems for cross-disciplinary scientific reasoning spanning gravitational-wave astronomy, cosmology, and effective field theories. We demonstrate that performance is enhanced by equipping agents with domain-specific skills, and apply our system to forecast how next-generation gravitational-wave observatories can distinguish between effective field theory models and probe cosmic reionization histories.

Flagship spectroscopic surveys such as DESI, LAMOST, SDSS, and PFS are accumulating an unprecedented dataset of stars spanning the Milky Way and the broader Local Group. In near-field cosmology, these ancient stars act as time capsules: their chemical abundances, particularly the [α/Fe] vs. [Fe/H] diagram, encode the star formation history and hierarchical assembly of our Galaxy. Maximizing the scientific yield of these heterogeneous datasets is a major challenge, as traditional stellar parameter pipelines often struggle at low-to-medium spectral resolution and suffer from domain mismatch across surveys.

I will present data-driven ML frameworks to bridge this cross-survey divide. First, simple MLPs pre-trained on LAMOST adapt rapidly to DESI via few-shot transfer learning, recovering the Galactic thin-thick disk chemical bimodality smeared out in the classical DESI pipeline. Next, I introduce SpecCLIP, a spectral foundation model using contrastive learning with auxiliary decoders to align LAMOST and Gaia XP spectra into a unified embedding space while preserving modality-specific information. SpecCLIP delivers competitive stellar parameter estimation and enables cross-survey spectral retrieval and prediction. I will also discuss where foundation model embeddings offer advantages over direct spectral inputs for cross-survey transfer. Together, these approaches offer a practical path toward extracting the chemical fossil record at scale.

Can AI help us unlock the longstanding questions in science? Can it ask new and, perhaps, more relevant questions? Can we trust it to do so? Can we extract physical information from vasts datasets, even if we don’t know what we are looking for?  In this talk I will discuss the promise and challenges of cutting-edge AI models, and the need to focus on how to define and achieve “realism” from a physical rather than purely perceptual perspective. I will present recent progress on physical, probabilistic, and generative models for astronomy. I will also discuss the interplay of classical physics models, ML and agentic AI.

Cosmology is entering an era in which inference can be performed directly from maps and fields, using simulation-based inference (SBI) across both large-scale structure and CMB surveys, offering a way to trace the underlying rhythms of the cosmos across multiple probes and scales. 

I will begin by motivating field-level inference as an optimal approach to extracting cosmological information from cosmic structure and reconstructing the initial conditions of the Universe. I will present and benchmark a range of field-level methods, from differentiable forward modeling to machine-learning approaches (including diffusion models and beyond), and show the significant gains they deliver for DESI BAO. I will then present an SBI pipeline for large-scale CMB B modes in current ground-based experiments, showing how one can marginalize over complex Galactic foregrounds and improve constraints on the tensor-to-scalar ratio, r. 

I will conclude by presenting novel methods to make SBI robust, interpretable, and efficient, and by outlining its broader role in next-generation cosmology, including simulations being developed within the Simons Observatory ecosystem to enable SBI for multi-probe cross-correlations.

Modelling the connection between galaxy and dark matter halo properties is a fundamental task in both galaxy evolution and cosmology. The modern use of Graph Neural Networks has allowed a deeper exploration of these relationships on both a galaxy-by-galaxy and population level, demonstrating that galaxies, halos, and their relationships are greatly influenced by both their detailed spatial environments and temporal assembly history. However, it is unclear to what extent these two modalities provide distinct or overlapping constraints. In this talk, I demonstrate a full equivalence in the impact of halo environments and assembly histories on a broad range of baryonic galaxy properties. This result holds when simulating galaxy properties using both full magnetohydrodynamic codes or semi-analytic models. To achieve the full equivalence, the environment cannot be expressed as spherically averaged density shells, but must be encoded as a geometric graph, linking halos on sufficiently large spatial scales. I furthermore measure the linking lengths that give optimal predictions for each galaxy property, and show that these are directly related to the typical extent of all halo progenitors. The equivalence offers new insights into making accurate effective models of galaxy properties constrained by observables, as well as a tantalizing opportunities for measuring the assembly histories of host halos in the real Universe.

We extend our previous work for cosmological simulations to incorporate baryonic hydrodynamics. We present a two-stage deep learning model that reconstructs high-resolution (HR) baryonic fields from low-resolution (LR) hydrodynamical simulations at redshift z=3. The model takes as input an LR simulation along with the corresponding high-resolution initial conditions (HR-ICs). In the first stage, a stochastic super-resolution network generates HR baryonic fields from the LR input. In the second stage, a deterministic emulator refines these outputs using HR-ICs to recover small-scale structures, including displacement, velocity, internal energy, and gas/star classification.

Trained on paired LR and HR simulations generated with MP-Gadget, the model captures small-scale features of the intergalactic medium and reproduces Lyman-α forest observables down to the ∼100 kpc pressure-smoothing scale. We achieve sub-percent accuracy in overdensity, temperature, velocity, and optical depth fields, a mean relative error of 1.07% in the large-scale flux power spectrum, and better than 10% agreement in the flux probability distribution function. 

Importantly, the framework reduces computational cost by a factor of ∼260 compared to full hydrodynamical simulations at equivalent resolution. This approach provides a fast and accurate alternative for generating high-resolution cosmological fields, enabling efficient production of large-volume mock datasets for next-generation surveys.

We present a simulation-based inference pipeline to perform parameter estimation from Lyman-α forest data using Balanced Neural Ratio Estimation (BNRE) combined with Hamiltonian Monte Carlo (HMC). BNRE enables the learning of likelihood-to-evidence ratios while explicitly encouraging posterior distribution calibration. The learned ratio is then used as a surrogate likelihood within HMC, allowing for efficient posterior sampling.

To handle realistic observational conditions, including missing or masked data, we incorporate a transformer-based embedding architecture that operates directly on one-dimensional inputs. This enables the model to extract informative features from incomplete skewers without requiring imputation or relying on summary statistics.

Our implementation is built entirely in JAX, taking advantage of just-in-time compilation and automatic differentiation to achieve fast training and efficient gradient-based sampling required for HMC. We demonstrate the approach on preliminary datasets of Lyman-α optical depth and transmitted flux skewers, focusing on constraints related to the timing of cosmic reionization. Initial results show promising posterior recovery and calibration properties, even in the presence of partial observations.

Magnetic fields pervade the observable universe and radio interferometry offers a unique means of probing the magnetic fields along the line of sight through the modeling of Stokes parameters Q and U of the incident radiation. Our methodology is  an alternative to utilizing Faraday synthesis which has been shown to be comparable for deriving the intervening components of the Faraday depth spectrum. Fitting to the stokes spectra in Q and U particularly through Bayesian inference or nested sampling  was effective but slow. We present vroom-sbi where we utilize the methods of simulation based inference, particularly, neural posterior estimation to speed up inference for QU fitting. Our results are comparable to Faraday synthesis and QU fitting but provide a speed up of 100x over the current state of the art such as RMTools that utilize nested sampling. We provide an open code repository and tools along with the trained models via huggingface for a plethora of models spanning a majority of the use cases in radio astronomy.

Image reconstruction in radio interferometry is a classic ill-posed inverse problem: recovering a continuous sky brightness distribution from sparse Fourier (uv-plane) samples. While the standard CLEAN algorithm is robust for point sources, it often introduces artifacts when imaging extended, diffuse structures. Regularized Maximum Likelihood (RML) methods offer an alternative but face significant computational overhead and tuning challenges as target resolutions increase.  

In this talk, I present a framework that overcomes these limitations by modeling the sky brightness as a continuous neural network. Unlike traditional "black box" deep learning, our approach is self-supervised, optimizing the network to fit the visibility data of a single observation directly. By mapping 2D sky coordinates to intensity values, the network functions as a resolution-independent representation rather than a fixed pixel grid. This architecture captures large-scale structures and fine details simultaneously, surpassing CLEAN with double the resolution and four times the fidelity. I will demonstrate the method’s performance on both synthetic tests and real ALMA datasets, proposing a new paradigm for high-fidelity interferometric imaging.

Very Large Array (VLA) observations between 16-50 GHz are limited by constraints in wind speed, cloud cover, and interferometric phase stability. Wind speed and cloud properties can be forecasted directly using the products provided by numerical weather prediction models, but there is no direct way to forecast the phase stability. To address this gap we present a machine learning model to forecast the root-mean-square (RMS) phase from a combination of (1) the measured RMS phase up to the start of the forecast window and (2) forecasted meteorological properties (e.g., surface wind speed, lifted index, convective available potential energy). We train a gradient-boosting decision tree using the LightGBM software library on seven years of historical forecast analysis data and measured RMS phase in a 6-to-1 train/test split. Compared to a naive seasonal model as a baseline, the LightGBM model reduces the RMS error in the test dataset by a factor of 2.1 with a median absolute error of approximately 1 to 2 degrees when excluding monsoon season conditions. The model is currently in operational use at the VLA and being used improve the scheduling efficiency of high-frequency observations. In future work we aim to better generalize the model to other sites in order to support the Very Long Baseline Array and the Next-Generation VLA.

Modern radio interferometry telescopes generate immense volumes of sparse, high-dimensional, and highly irregular data, making it difficult to develop efficient amortized machine-learning models with standard architectures. A powerful solution is to adopt a better data representation: raw radio interferometric visibilities can be transformed onto a regular grid in Fourier space. We present VisCube, a visibility-space gridding and uncertainty quantification framework for radio interferometry. Gridding data with VisCube serves several purposes: the data is transformed into a compact representation, the resulting grid is compatible with existing image-based algorithms and architectures, and the output provides separate estimates of both signal and uncertainty. By treating signal and uncertainty as distinct outputs, VisCube enables probabilistic machine-learning and Bayesian inference pipelines directly on interferometric data. We highlight several ongoing projects that leverage VisCube, ranging from neural imaging pipelines to state-of-the-art supermassive black hole mass measurements. 

Scheduling telescope observations is traditionally accomplished with bespoke heuristic algorithms, which are increasingly outpaced by the operational complexity of modern astronomical surveys. These static heuristics often struggle to balance conflicting scientific priorities, long-term planning, and stochastic environmental conditions. This lack of adaptive tools increases the human cost of manual scheduling and can lead to systematic inefficiencies in survey execution.

In this talk, I will present our research on intelligent scheduling using reinforcement learning. We initially utilize behavior cloning on archival schedules to develop an agent capable of mimicking expert human/heuristic scheduling without explicit rule programming. By framing survey operations as a high-dimensional optimization problem, we establish a framework to tackle fundamental challenges in multi-objective optimization and the management of conflicting rewards across competing timescales. We have implemented this approach on the NSF Victor M Blanco 4-meter Telescope, one of the most impactful ground-based telescopes in operation.

Finally, I will discuss a vision for integrated autonomous observatories. We envision a future where our methods generalize across wavelengths and facilities, eventually guiding everything from proposal design to real-time target selection.

I will discuss the impact that recent developments in AI have for the running of the IDIES data center and in particular the SciServer science platform. Foundation models and natural-language interfaces are reshaping what users expect from a science platform and how we provision data and compute behind it.

On the data side, SDSS DR20 is being collocated with major external spectral collections on our 10+PB CEPH cluster and prepared for modern ML algorithms. We are exploring Parquet and Zarr storage with Croissant and other (meta)data formats, and extending SDSS cross-matching so embedding-level similarity search can sit alongside positional and spectral matches.

On the compute side, GenAI now runs on SciServer alongside notebook and SQL workloads. We are disseminating foundation models for inference through vector databases and adding natural-language interfaces over our relational stores. A campus-wide WEKA file system will link SciServer to larger GPU and HPC resources, so heavy training no longer requires moving data off-platform.

These capabilities are not unique to astronomy. SciServer also hosts data from turbulence and cosmological simulations, oceanography, microscopy, and materials science. Partnerships with the JHU Data Science and AI Institute (DSAI), the Scientific Software Engineering Center (SSEC), and RSE support let advances cross domain boundaries.

Over the coming decade, time-domain survey instruments will discover millions of extragalactic transient candidates. Photometrically classifying these transients will be necessary; however, most current classification efforts are ill-equipped to handle the large volumes of transients discovered at high redshifts, where the rest-frame ultraviolet emission is redshifted into the observer-frame optical and infrared. To address this shortcoming, we have created new multi-dimensional light curve and spectral templates of different spectroscopic classes of extragalactic transients, all containing ultraviolet observations, with a particular focus on core-collapse supernovae. Our novel methodology applies Gaussian Process Regression (GPR) modeling to observations of transients from the ultraviolet to the infrared across a range of phases. GPR is a data-driven machine-learning framework which naturally interpolates between observations and provides robust uncertainty estimates. This enables a forecasted light curve or spectrum at arbitrary wavelength or phase. We will demonstrate the process of generating these models and describe our open-source codebase, made publicly available to encourage community engagement. We will also demonstrate how these models can be used to generate synthetic datasets to train photometric classification algorithms. Finally, we will conclude by describing a number of physical inferences made possible by these model templates.

LSST is detecting an unprecedented number of transients each night, far exceeding the classification capacity of all existing spectroscopic resources. As a result, robust photometric classifications is essential both for assembling complete samples of different transient subtypes and for identifying events that warrant spectroscopic follow-up.

In this talk, we introduce a family of highly performant, real-time hierarchical classifiers for ZTF and LSST alert streams. These models incorporate images, multi-band light curves, and contextual metadata, to deliver reliable, high-level classifications within seconds of the first alert. Our hierarchical approach is unique in its ability to vary the granularity of classification based on the available data, making it possible, for the first time, to produce high-level classifications at early times, and refine those classifications as more observations become available.

The Lyman-alpha (Lyα) forest is the most powerful probe of cosmic expansion at redshifts 2 < z < 4. Cosmological parameters can be measured from its two-point correlation function through features such as the BAO scale and the Alcock-Paczyński (AP) effect. However, current analyses are limited by the quasar continuum fitting process, which removes information on large scales and distorts the correlation function on all scales. To address this, I developed the Lyα Continuum Analysis Network (LyCAN), a convolutional neural network that predicts the forest continuum using only longer-wavelength features. I will present LyCAN results on DESI DR1 data, providing the most precise measurement to date of the effective optical depth evolution. Additionally, Lyα forest 3D correlation analyses are limited by covariance matrix estimation, as the large number of correlation function bins relative to the smaller number of independent subsamples leads to significant noise in the measured covariance. Current analyses mitigate this with an ad-hoc smoothing procedure, but this does not fully preserve the covariance structure. I will present recent work developing a method to "denoise" the measured covariance matrix, resulting in more robust cosmological constraints and enabling a future LyCAN-based full-shape analysis. Such an analysis could improve AP parameter constraints by up to ~15%, strengthening the forest's ability to distinguish between competing cosmological models at high redshift.

Observations reveal a strong correlation between the masses of SMBHs and the properties of their host galaxies, suggesting a coevolutionary relationship between BHs and galaxies. To understand the underlying physical mechanisms originating these relations, it is essential to study the statistical properties of AGNs, which are driven by BH fueling and feedback. Galaxy mergers are interesting events as gravitational torques during mergers drive gas towards galactic nuclei, which can enhance AGN activities. Dual AGN activity is an electromagnetic precursor of SMBH mergers that produce gravitational waves. Thus, we investigate and characterize the single and dual AGN activities using isolated and merging galaxy simulations called SMUGGLE. Multiphase ISM in SMUGGLE simulations yields highly variable accretion rates with short duty cycles. The role of stochastic, nonuniform time-series analysis becomes imperative in this task. We also utilize interpolation and function fitting methods to analyze and characterize the duty cycles. We found the mean active phase timescales to be in the order of ~0.1-1 Myr and quantified dual AGN lifetime fractions in various merger scenarios. Differences in black hole masses, galaxy morphologies, and wind speeds suggest a significant impact on (dual) AGN activities. This work presents a method for modeling realistic (dual) AGN populations in SAMs and simulations, and thus for characterizing BH populations using constraints from EM and GW observations.

Data gathered by an interferometer requires substantial processing before astronomers can extract the scientific information from it. The selection and sequencing of calibration and analysis actions is a complex decision-making task. This decision process depends on the expertise of astronomers, who consider data characteristics, instrument knowledge, compute cost, and best practices to guide each choice. We apply reinforcement learning (RL) to this task, where an agent can autonomously explore and identify optimal decisions based on an objective function with metrics that quantify best practices. By framing data processing as a pathfinding and cost minimization problem, we can use this data-driven approach to learn effective sequences and settings of algorithms for data processing. This has implications for improving the accuracy and efficiency of high-throughput automated interferometric data processing.

Why RL? RL is ideal for tasks where hard-coding an action for every situation isn't feasible. Unlike supervised learning, RL doesn’t require labeled data; instead, it relies on a feedback loop from an objective function to improve its decision-making.

The Very Large Array Sky Survey (VLASS) is a large, multi-epoch, S-band (2–4 GHz), VLA B-configuration radio survey that delivers high-resolution (~2″) imaging over a wide area of the sky, making it well suited for discovering unusual radio sources. In this work, I will present an automated VLASS anomaly-finding pipeline designed to identify extended, non-Gaussian, and morphologically complex radio sources, including radio galaxies and other diffuse or structured objects, while rejecting compact point-like sources. The pipeline processes VLASS Quick Look (QL) median stacked maps, measures source and shape properties, assigns anomaly scores, generates cutouts, and produces an HTML-based gallery and candidate catalog for inspection and follow-up. By enabling systematic detection of rare and extended radio morphologies at survey scale, this framework supports the construction of new catalogs of anomalous radio sources from VLASS data. I will also make use of the VLASS Single Epoch (SE) maps which are self-calibrated images and have better signal-to-noise ratio. Future work will focus on machine-learning-based classification, improved ranking of astrophysical candidates versus artifacts, and automated cross-matching with optical data to identify likely counterparts and better characterize the nature of these sources. 

Modern radio astronomy pipelines produce a large amount of bandpass calibration data, and platforming anomalies (which manifest as lower-intensity readings across a frequency channel interval) are a common failure mode that can impact downstream calibration if not reliably detected. These anomalies are characterized by structured, contiguous deviations embedded within smooth, correlated spectral backgrounds. In this work, we present a solution to this anomaly-detection problem in bandpass spectra by developing a new scan statistics algorithm based on Nadaraya-Watson kernel regression (NWKR). While one could approach this as a high-dimensional estimation task, these generic embeddings often fail to capture relationships between coordinates, making it difficult to identify interval-structured anomalies. Our approach directly scans for this interval structure. Each candidate interval is evaluated by comparing a global regression fit to separate fits inside and outside the interval. While several regression models can be used, we find that NWKR works effectively, and through algorithmic insights, we can make it very efficient.  

We evaluate the proposed algorithms on both synthetic datasets designed to mimic realistic spectral structure and on bandpass data from the multiple stages of the quality assurance pipelines. The results show that NWKR-based scan statistics achieve high localization accuracy for platforming anomalies and significantly outperform baseline methods.

We present an architecture for LLM-driven radio interferometric data reduction built on a strict separation between measurement and reasoning. A Model Context Protocol (MCP) layer wraps CASA, exposing each operation on a Measurement Set — metadata queries, instrument geometry, flagging, calibration — as an independent tool returning structured data with explicit completeness and provenance annotations. No tool interprets its output or chains to another. Reasoning is supplied separately: version-controlled domain documents encoding interferometric expertise are injected stage-by-stage into a small (4B-parameter) language model running on a single consumer GPU. An orchestration layer routes the model through a workflow — inspection, human checkpoints with structured decision capture, and iterative calibration tolerant of long-running CASA jobs. The model is stateless; all durable context lives in a persistent store, making every decision traceable. Human oversight is structural: at defined checkpoints the model surfaces findings as structured questions; the astronomer's answers propagate forward as configuration the model must respect. We argue that for domain-specific scientific workflows, protocol design and knowledge curation yield better returns than scaling model parameters. We demonstrate end-to-end inspection and calibration of VLA continuum data and discuss extension to MeerKAT and uGMRT.

Modern data analyses frequently encounter settings where samples of variables are contaminated by measurement error. Ignoring measurement noise can degrade statistical inference and existing approaches can often be computationally inefficient in high dimensions or make strong assumptions about the form of the prior distribution. Recent advances, particularly those based on Maximum Mean Discrepancy (MMD), have enabled flexible, distribution-free inference, which typically assume precise data and overlook contamination by measurement error. In this work, we introduce a novel framework for deconvolution based on convolutional MMD (convMMD) which compares distributions after noise convolution and retains metric validity under standard kernel conditions. We show an equivalence between estimation under noise and kernel smoothing, and provide robustness guarantees for mild misspecification of the noise distribution. Leveraging these insights, we introduce a convMMD-based empirical bayes estimator for inference and deconvolution with noisy, heteroscedastic observations. We establish its consistency and asymptotic normality, and provide an efficient implementation using stochastic gradient descent. We demonstrate the practical effectiveness of our approach through simulations and applications in astronomy.

Radio astronomy is entering a data-driven era with advances such as the ALMA Wideband Sensitivity Upgrade (WSU) and next-generation facilities: ngVLA and SKA. Traditional anomaly detection in radio data, particularly for bandpass calibration, relies on heuristic thresholds or manual inspection and is neither robust nor scalable to upcoming data volumes.

We present a self-supervised framework for bandpass quality assessment using a masked variational autoencoder (VAE) trained to learn the low-dimensional manifold of good (well-calibrated) solutions. The architecture comprises three CNN-based encoder layers and ingests dual-channel inputs: bandpass amplitude and atmospheric transmission. It is trained with a median absolute deviation (MAD)-weighted reconstruction loss to reduce sensitivity to noisy data and missing channels. The decoder is conditioned via FiLM modulation, and a conditional latent prior ensures atmosphere-driven variations are captured in the latent space rather than misclassified as anomalies. The model (~400K parameters) is trained on ~80K ALMA bandpass solutions across various observing bands, correlator setups, and observing conditions.

Anomalies are identified using robust statistics on reconstruction errors, separating bad calibration solutions from expected variability. The pipeline is lightweight and scalable to next-generation facilities. Future work will extend this method to additional modalities (e.g., phase) and higher-dimensional visibility data.

In this presentation I will discuss new concepts on how astronomy can better enable AI across large and diverse data sets. I will start with the value of working across data sets with conventional tools, and the impact that work has had to date. I will touch on the special and somewhat paradoxical place astronomy finds itself today in the context of modern data systems. I will discuss how unified data systems can enable AI in particular, and broaden the community working with astronomical data and point out places where, conversely,  AI can help with this data unification process. I’ll close by discussing some particular architectural ideas about how new data systems can best serve astronomers deploying AI, and soliciting your ideas of what you want in the AI-oriented astronomy data systems of the future.

IPAC is currently engaged in several efforts to understand the utility of AI, from the development and operation of pipelines, to the ingestion of literature data into archives and in assisting users to understand and extract data from archives.  In this talk I will present a summary of those efforts and also describe some of the programmatic issues we have encountered in implementing user-facing functionality incorporating AI.  These include funding models, navigating policies on specific vendors and dealing with AI-driven data harvesting.

Next-generation astronomical facilities, including the Rubin Observatory, Dark Energy Spectroscopy Instrument (DESI), and the Roman Space Telescope, are delivering multimodal data of unprecedented volume. Extracting scientific insights from these colossal datasets requires robust, automated computational frameworks. Foundation models, optimized via self-supervised learning on unlabelled data, offer a highly scalable solution to this impending data deluge.

This presentation introduces the AstroVision model for Imaging and Spectroscopy (AVIS), a novel foundation model designed specifically for astronomical research. AVIS was trained on 2 million spectra from DESI Data Release (DR) 2 with a wide variety of objects, generally divided into three main categories: stars, galaxies, and quasars. I will outline AVIS’s architecture and demonstrate its efficacy across key applications, such as redshift estimation and object classification. While traditional methodologies rely heavily on labor-intensive visual inspection, AVIS significantly reduces this human bottleneck while maintaining rigorous baseline accuracy. Finally, because interpretability is essential for scientific adoption, this talk explores the benchmarking and explainability techniques utilized to evaluate the model’s capabilities.

I will talk about our recent work building a new simulation suite, emulator and likelihood to make cosmological parameter constraints from the eBOSS and DESI Lyman-alpha forest, absorption from neutral hydrogen gas spread throughout the intergalactic medium. The PRIYA simulations are a large suite of cosmological hydrodynamic simulations, with 60 different cosmologies combining a large box, high resolution and a state of the art galaxy formation model. Using them we placed interesting constraints on the slope and amplitude of matter clustering at z=3, where AGN feedback effects are expected to be weak.

Initial condition (IC) reconstruction aims to infer the configurations of the primordial density perturbations that seeded the observed large-scale structure of the universe. In this talk, I will show how 3D diffusion models can reconstruct ICs while accounting for a comprehensive range of physical processes, from gravity and hydrodynamics to complex astrophysics. While traditional methods for IC reconstruction are often limited to quasi-linear simulators, due to the necessity of differentiability in Hamiltonian Monte Carlo (HMC) sampling, our fully amortized diffusion model can invert N-body and hydrodynamical codes without requiring MCMC sampling. This allows us to target smaller scales and amortize the inference procedure, providing comprehensive modeling of gas hydrodynamics and astrophysical feedback. Using a U-Net architecture trained on the IllustrisTNG CAMELS suite, we demonstrate this method’s robustness to physical priors by marginalizing over various cosmological and astrophysical models. Once trained, this approach is orders of magnitude faster than HMC, providing a scalable solution for the large volumes of data expected from modern galaxy surveys.

Observing from inside the Milky Way provides a unique opportunity to probe galaxy properties with a level of precision that cannot be achieved for more distant systems. Observations of the Milky Way can be used to constrain both the nature of dark matter and baryonic feedback uncertainties. A key component of such studies is the comparison of Milky Way observables with Milky Way-like systems simulated using varied dark matter and baryonic feedback models. However, it remains unclear what criteria are required for a simulated galaxy to be considered a true Milky Way analog. While matching fundamental properties such as halo mass is necessary, more subtle characteristics of the Milky Way have been found to play an important role. In this work, we investigate how the Milky Way’s surrounding environment impacts its satellite population. This work uses the Cold Dark Matter suite of 1024 Milky Way zoom-in simulations from the DREAMS Project. We augment this data by re-running the corresponding N-body suite with a larger volume at lower resolution to study the Milky Way’s environment out to a few megaparsecs. We train a generative emulator on these two suites to produce additional Milky Way galaxies across a range of masses and environments. Our results suggest that the Milky Way's environment, quantified by the distance to the nearest galaxy with a mass ≥ the Milky Way’s mass, does not impact the abundance, mass function, or radial distribution of subhalos.

As cosmological simulations reach unprecedented resolutions, disentangling the physical effects of baryonic feedback from the numerical artifacts of specific hydrodynamic codes remains a critical challenge. We present the DREAMS Dwarf Suite, a new set of high-resolution dwarf galaxy simulations (~$10^{10} M_{\odot}$) independently run across three distinct hydrodynamical codes: FIRE3, RAMSES, and ChaNGa. This work establishes the baseline properties of the suites, focusing on the stellar mass-halo mass relation, star formation histories, and metallicity of low-mass galaxies. By systematically varying sub-grid implementations, such as star formation prescriptions and supernova feedback efficiencies across different numerical architectures, we quantify the theoretical uncertainties inherent to modeling the faint universe. To quantify these uncertainties, we introduce a novel weighting scheme constrained by the observed scaling relations [arXiv:2512.00148]. By applying these pseudo-posterior constraints, we demonstrate that standard single-model tuning misses complex parameter interdependencies. Crucially for the intersection of AI and astrophysics, these presentation suites serve as a robust, multi-code training dataset. By moving beyond single-model simulations, the DREAMS Dwarf Suites provide the necessary high-dimensional data to train the next generation of generative emulators and SBI models, enabling researchers to robustly marginalize over numerical uncertainties.

Over the last decade, large hydrodynamic simulations of galaxy formation have become the state-of-the art when it comes to theory predictions about the large scale structure of the Universe, galaxy clustering, the formation and evolution of galaxies, and recently also strong lensing. To understand the formation and evolution of galaxies across mass scales and in realistic environments it is desirable to use large volume simulations, necessitating the use of effective models for, e.g., star formation, stellar feedback, black hole accretion, and blackhole feedback. In the recent past, a lot of progress has been made in the field of smaller scale, full physics simulations of both blackhole and stellar physics. Within the Learning the Universe collaboration, our focus has been on leveraging these results to develop new, physics-informed subgrid models for cosmological simulations. In this talk, I will outline our efforts to combine these subgrid models into a new galaxy formation model with the help of simulation based inference techniques. In particular, I will showcase the calibration of a new, torque-limited black hole accretion model as a case study, before outlining the full calibration effort, which relies on several suites of zoom simulations featuring different combinations of subgrid models, performed with the aim of finding the model most compatible with benchmark observations.

Machine learning is increasingly used to connect galaxy observables to galaxy physical properties, but its scientific value depends on interpretability. I will present ongoing work applying explainable AI to machine-learning models of reionization-era emission-line galaxies, using extreme emission-line galaxies (EELGs) and Lyα emitters as motivating science cases. Using JWST samples, including EELGs identified through NIRCam F410M excess in the JWST public fields, we train tree-based regression models to predict emission-line strength from inferred galaxy properties such as recent star-formation burstiness, stellar mass, UV luminosity, UV slope, dust attenuation, metallicity, and star-formation rate. We then use SHAP to quantify which physical parameters drive the predictions across the galaxy population and for individual galaxies. Our current results suggest that recent burstiness is the strongest driver of extreme Hβ+[O III] equivalent widths, with sSFR, dust attenuation, stellar mass, and metallicity also contributing. This work is a first step toward a broader interpretable-ML framework for linking nebular emission to bursty star formation, ionizing photon production, and environment during the epoch of reionization.

The Simons Collaboration on Learning the Universe (LtU) is a large collaborative effort to develop and apply machine learning and simulation-based inference methods to extract galaxy formation and cosmological information from astronomical observations. In this talk, I will present an overview of recent progress across the collaboration, spanning the development of new physically motivated simulation models, the generation of large and diverse training sets, and the deployment of scalable inference pipelines capable of meeting the demands of next-generation surveys. I will then discuss a small number of these efforts in greater depth, highlighting advances in field-level emulation, diffusion-based generative modeling, and robust neural inference.

To handle the deluge of data alerts expected nightly by the Vera C. Rubin Observatory's LSST, we introduce SELDON (Supernova Explosions Learned by Deep ODE Networks), a custom time-domain AI foundation model designed for millisecond-scale inference. SELDON utilizes a continuous-time VAE combining a masked GRU-ODE encoder with a latent neural ODE propagator to natively process sparse, gappy, heteroscedastic, nonstationary, and irregularly sampled multisurvey multichannel time-series data. An interpretable basis decoder maps these latents into physically meaningful, scale- and time-invariant basis functions. SELDON enables early-phase forecasting, predicting critical evolution metrics before the light curve peak when data are extremely limited. We calibrate uncertainty via a latent Schrödinger bridge combining optimal transport with latent diffusion.

To incorporate upcoming massive fiber-fed spectroscopic surveys, we extend SELDON into a multimodal framework. We introduce a customized transformer using transfer learning to map rich, simulated spectroscopic data SASSAFRAS to imbalanced WISeREP observations. To overcome limited quality control across heterogeneous instruments, a novel multi-resolution analysis effectively denoises public spectra before probabilistic classification. Fusing continuous-time photometry with deep spectroscopic transformers, SELDON establishes a real-time AI blueprint for downstream alert filtering, anomaly detection, and automated follow-up.

From plasma flow to radiative transport, advection-dominated phenomena play a central role in astrophysical systems, where sharp fronts and coherent structures are transported over time in the spatial domain. Traditional reduced-order modeling (ROM) uses global bases fixed in time; thus, efficiently capturing these features remains a major computational challenge. We introduce a semi-Lagrangian (SL) ROM framework that addresses the limitations by advecting time-dependent local bases via transport maps. It exploits the underlying physics by tracing characteristics, while expressing the dynamics on the Eulerian grid via interpolation. We propose two distinct algorithms: one using nearest neighbors with velocity splitting to separate the transport and correction terms, and another using deterministic sketching via pivoted QR. We demonstrate the framework on 1D parametrized linear advection equation with high variability in velocity fields and initial conditions, and later extend to 2D settings. We have conducted a detailed complexity analysis showing substantial speedups over FOM, along with error bounds to ensure a trustworthy ROM. Numerical results confirm that SLROM using local bases can overcome the limitations of global-basis approaches and capture key features under high variability in the advection field. Hence, it is a numerically efficient, scalable, and physics-informed approach to modeling transport-driven systems in astrophysics.

Advisors: George Biros, Omar Ghattas

Cosmological simulations are essential for survey interpretation but are limited by a volume–resolution trade‑off. ML super‑resolution (SR) can add small‑scale structure to low‑resolution (LR) runs, but SR is an inverse problem: one LR field maps to many plausible HR realizations. I will present a two‑stage pipeline that turns stochastic SR into a controllable emulator. A time‑conditioned SR model in Lagrangian phase space increases mass resolution while preserving large‑scale modes and matching standard structure statistics. A second, deterministic emulator uses LR outputs plus information tied to HR initial conditions to recover the specific HR realization, improving structure alignment and Fourier‑mode cross‑correlation with HR. I will also show an extension to hydrodynamical simulations for Lyα‑forest applications, achieving high‑fidelity sightline predictions with large speed‑ups over full hydro runs.

Constructing surrogate models for computationally expensive functions is an important challenge in science and engineering. Surrogates are used to accelerate important computations in design, control, uncertainty quantification and optimization. In the context of predicting solutions to high-dimensional ODEs, many approaches can be challenging to use in practice due to slow inference times and lack of scaling to high dimensions. Here, we present a new surrogate algorithm that is designed to predict the flowmap of ODEs. Our algorithm is based on nearest neighbor interpolation, which makes our algorithm robust and allows for built-in accuracy guarantees. We also include higher-order information in the form of gradients. We use adaptive clustering so that our surrogate is accurate according to user-specified tolerance. Finally, we use higher-order information in the form of Taylor expansions, which we find greatly improves accuracy of our surrogate. The target application is accelerating the astrochemistry component of star formation simulations. We provide an open-source parallel implementation of the algorithm for both CPU and GPU architectures. We compare our algorithm with state-of-the-art approaches, including standard nearest neighbor interpolation and a deep neural network. The results indicate that the accuracy is comparable to these approaches, is easier to train, and scales well to high-dimensions. Co-authors: Prajwal Prathiksh, Skandan Subramanian, George Biros

Standard projection-based model reduction for dynamical systems incurs closure error as it only accounts for instantaneous dependence on the resolved state. From the Mori-Zwanzig (MZ) perspective, projecting the full dynamics onto a low-dimensional resolved subspace induces additional noise and memory terms arising from the dynamics of the unresolved component in the orthogonal complement. The memory term makes the resolved dynamics explicitly history dependent. In this work, based on the MZ identity, we derive exact, history-enriched models for the resolved dynamics of linear dynamical systems and formulate inverse problems to learn model operators from discrete snapshot data via least-squares regression. We propose a greedy time-marching scheme to solve the inverse problems efficiently and analyze operator identifiability under full and partial data availability. For full data, we show that, under mild assumptions, the operators are identifiable even when the full-state dynamics are governed by a general time-varying linear operator, whereas with partial data the inverse problem has a unique solution only when the full-state operator is time-invariant. To address the resulting non-uniqueness, we introduce a time-smoothing Tikhonov regularization. Numerical results demonstrate that the operators can be faithfully reconstructed from both full and partial observation data and that the learned history-enriched MZ models yield accurate trajectories of the resolved state.

Accurate modeling of baryonic physics remains a major challenge for precision cosmology due to our incomplete understanding of complex subgrid processes, like star formation and feedback from supernovae and active galactic nuclei below ~10 Mpc scales. This uncertainty leads to different hydrodynamical simulation suites to implement fundamentally different prescriptions for these unresolved physics. We introduce a machine learning framework that learns continuous representations of baryonic feedback across multiple simulation suites, to enable interpolation between different physical implementations while providing robust uncertainty quantification. Our approach addresses the key challenge of marginalizing over theoretical uncertainties represented by various simulators while simultaneously constraining the underlying baryonic physics from observations. We frame this as learning a shared continuous latent representation of the physics implemented across different simulators, allowing us to both marginalize over and constrain a continuous baryonic parameter space. Using CAMELS, we demonstrate our method on several baryonic fields including stellar mass, gas density, temperature, and pressure fields. This framework provides a path toward more robust cosmological inference by properly accounting for theoretical uncertainties in baryonic modeling while extracting maximum information about the underlying physical processes from current and future surveys.

Over the last decade, cosmological simulations have become an invaluable tool for understanding how galaxies form and evolve.

Yet, despite their successes, different flagship models often disagree on key observables, revealing that the achievements of the current generation lay on uncertain foundations.
The DREAMS (DaRk mattEr and Astrophysics with Machine learning and Simulations) simulation efforts have a primary objective of improving these foundations through systematic variations in dark matter physics, cosmology, and astrophysics. 

These variations, coupled to machine learning algorithms, allow us to more effectively characterize our simulation models, understand key uncertainties in them, and more efficiently plan future efforts.

This talk will highlight how our new Halo Mass Varied suite is able to recreate full box simulation statistics with a modest number of targeted zoom-in simulations.
Moreover, with our new Resolution Varied Suite, we can extend this even further by providing an effective ``super-resolution'' mapping from low resolution outputs to high resolution outputs.

Together, these efforts represent a speed-up of several orders of magnitude, offering a scalable path for incorporating model uncertainty into the next generation of galaxy simulations.

We present a conditional diffusion model framework to infer two-dimensional dark matter density profiles directly from realistic deep imaging data. Our approach is trained on 1024 Milky Way–mass halos from the DREAMS simulation suite, spanning variations in cosmology and baryonic feedback prescriptions. Mock Euclid VIS images are generated using the forward-modeling pipeline, Synthesizer. The model learns a stochastic bridge between observed baryonic light distributions and underlying dark matter density maps. For held-out halos, we recover 2D dark matter density profiles with a typical accuracy of $\sim0.15$ dex across resolved spatial scales. The stochastic sampling procedure enables uncertainty estimates that reflect both halo-to-halo variance and baryon–dark matter degeneracies. We further validate the framework against redshift dependence and domain shifts across simulation variations. Our results demonstrate that diffusion-based generative models offer a flexible, uncertainty-aware framework for translating forthcoming data from large imaging surveys (Euclid, Roman, Rubin, etc.) into spatially resolved dark matter constraints, opening new opportunities for statistical studies of galaxy–halo connections in large photometric surveys.

Modern cosmological surveys are mapping the universe at an unprecedented scale and resolution, offering a stringent testbed to probe the physics of dark matter, dark energy, and structure formation. Machine Learning (ML) is transforming how we process these massive datasets, capturing rich, complex information that traditional methods discard. However, despite their statistical power, ML methods are rarely applied to real observational data due to fundamental challenges in simulation realism, numerical scaling, and prediction stability.

This talk presents how the Learning the Universe (LtU) collaboration is productionizing simulation-based inference (SBI) for robust parameter estimation on modern surveys. Using our flagship cosmological analysis of the SDSS CMASS spectroscopic galaxy clustering as an empirical testbed, I demonstrate how we leverage emulators to scale high-resolution simulations to cosmological volumes while validating their physical fidelity. I show that training SBI at scale with these  carefully curated datasets enables maximal information extraction from observables while ensuring robustly calibrated uncertainty. Finally, I introduce LtU-ILI, our public inference code designed to democratize these production-ready tools for both early-career researchers and large survey collaborations.

Dark matter direct detection experiments require information about the local dark matter speed distribution to produce constraints on dark matter candidates, or infer their properties in the event of a discovery. I will discuss how the uncertainty in the dark matter speed distribution near the Sun is affected by baryonic feedback, halo-to-halo variance, and halo mass. I will utilize the statistical power of the new DREAMS Cold Dark Matter simulation suite, which is comprised of 1024 zoom-in Milky Way-mass halos with varied initial conditions as well as cosmological and astrophysical parameters. Applying a normalizing flows emulator to these simulations, the uncertainty in the local dark matter speed distribution is dominated by halo-to-halo variance and, to a lesser extent, uncertainty in host halo mass. Uncertainties in supernova and black hole feedback (from the IllustrisTNG model in this case) are negligible in comparison. Using the DREAMS suite, I will present a state-of-the-art prediction for the dark matter speed distribution in the Milky Way. Although the Standard Halo Model is contained within the uncertainty of this prediction, individual galaxies may have distributions that differ from it. Lastly, I will discuss applying the DREAMS results to the XENON1T experiment and show that the astrophysical uncertainties are comparable to the experimental ones, solidifying previous results obtained with a smaller sample of simulated Milky Way-mass halos.

Self-interacting dark matter (SIDM) spans diverse particle-physics models: from contact interactions to light-mediator scenarios with velocity- and angle-dependent cross sections and complex dark sectors with multiple states, inelastic transitions and dissipative processes. Representative classes include thermal SIDM with light mediators, resonant SIDM, vector SIDM, asymmetric SIDM, multi-state inelastic models, composite SIDM (dark baryons, mesons, atoms), dissipative SIDM and decaying DM.

We have developed a novel implemention of SIDM in the moving-mesh code AREPO-2, together with a framework that maps particle-physics models onto simulation inputs. The modular implementation handles velocity-dependent and anisotropic cross sections, inelastic channels, dissipation and decays. We have tested these algorithms across a range of idealised and cosmological setups, and assessed their performance and scalability in isolated core-collapse simulations and in cosmological boxes, both DM-only and with baryons. Except during the late stages of gravothermal core collapse, SIDM runs incur only modest overhead relative to the corresponding CDM simulations, and are substantially faster than the previous SIDM implementation in AREPO-1.

Looking ahead, we aim to use this pipeline to run simulation suites and to develop machine-learning methods for inferring DM properties from cosmic structures, exploring how far observations can constrain the particle nature of DM.

The Vera C. Rubin Observatory was conceived in the Big Data era around what we would now call AI, long before LLMs. Built to robotically image the entire southern hemisphere every few nights with a 3.2-gigapixel camera, Rubin would produce a time-lapse of the night sky, revealing moving asteroids, pulsing stars, supernovae, and rare transients that you only catch if you're always watching.  Its automated observation scheduling, image processing algorithms, and real-time alert distribution were designed on the premise that scientific discovery would emerge from machine learning on petascale data releases rather than targeted observations.

As Rubin starts its Legacy Survey of Space and Time (LSST), the AI landscape today looks quite different from what we envisioned.  For example,  foundation models are changing how scientists expect to interact with archives and compute platforms.  I will discuss the lessons learned from preparing the observatory for science operations and how today's AI landscape is reshaping our plans to produce and serve optical astronomy's largest data set.

What would happen if you applied a model trained entirely outside astronomy to time series of bright AGNs — no feature engineering, no domain adaptation, just the data? Using frozen foundation models as feature extractors — with a lightweight head trained where needed — we evaluate tasks ranging from predicting observables across bands to physical parameter inference. We use archival data spanning radio to gamma rays, and go beyond lightcurves by including morphological metrics. These models prove competitive with expertly hand-crafted variability features even for challenging multi-band and highly unevenly sampled data. Carefully designed classical features can still outperform them on specific tasks, but they fundamentally require domain expertise. We further find that the learned embedding space carries real astrophysical meaning: a map of blazar variability that no one designed, yet proves practical to navigate. The broader time series foundational models ecosystem, built without astronomy in mind, turns out to already be useful for AGN science — just bring the data.

The chemical abundances of stars are tightly coupled to the underlying physics of galaxy evolution. However, chemical evolution is a challenging, underdetermined inverse problem, subject to strong degeneracies (especially between star formation and nucleosynthesis) and substantial but often unquantified uncertainties. Posing chemical evolution as a population-level phenomenon can break these degeneracies, enabling robust constraints on metal production and the baryon cycle in galaxies. In this talk, I will present a novel hierarchical Bayesian model of chemical evolution that isolates environmentally-dependent processes like star formation and galactic outflows from universal nuclear and stellar physics by conditioning directly on observed star formation histories and metallicity distributions. Further, by jointly modelling multiple galaxies in a partially-pooled model, this model naturally explores environmental, mass, and metallicity dependencies of processes in galaxy evolution. I will discuss its application to local dwarf spheroidal galaxies, which revealed evidence of a metallicity-dependent rate and delay-time distribution of Type Ia supernovae.

The Long Wavelength Array (LWA) is a distributed low frequency radio telescope with stations in Texas, New Mexico, Arizona, and California. The LWA relies on a high degree of automation and station autonomy to help make management of the stations easier for a small team. At the core of this are the HAL systems, “software defined operators”, and a collection of utilities that automate the observing, archiving, and data delivery process. Although these tools have greatly improved the operational efficiency of the telescope, they still rely on a human in the loop to diagnose and resolve some types of problems or error conditions.

To determine if we can further reduce the time that people spend diagnosing and fixing problems, we have been experimenting with AI/ML tools called Alan and Oarfish. Alan is a chatbot that uses retrieval-augmented generation to help monitor the stations, diagnose problems, and provide students with help analyzing LWA data. Oarfish is a convolutional neural network that provides real-time classification of all-sky images for interference and station health monitoring. In this talk I discuss how these are implemented and touch on how these tools fit in with the LWA Swarm concept where multiple institutions build and operate LWA stations. Integration of these AI tools into the Swarm has the potential to lower the barriers of entry for universities that do not have a background in instrumentation or low frequency radio astronomy.

During interferometric radio observations, the bandpass calibration is a procedure to normalize atmospheric transmission by recording antenna responses of bandpass calibrators, referred to as calibration solutions. The reliability of these solutions is often assessed with human inspection and expert-crafted heuristics implemented in the observatory pipeline. Alternatively, treating the problem as a supervised classification task, we use a Gradient Boosting model to classify anomalies in the amplitude solutions of ALMA bandpass calibrations. The model uses features derived from kernel-based scan statistics to characterize deviating intervals and is trained on a dataset of bandpass solutions from ALMA Cycle 9 observations. By optimizing predictive performance, the ML model achieves a recall of 0.92 and an F2 score of 0.86, with significantly lower false positive and false negative rates than the bespoke heuristic used in the ALMA pipeline. By reducing missed anomalies and false classifications, this approach improves flagging accuracy and calibration solution quality, providing a scalable supplement to support observatory operations.

Stage-IV dark energy wide-field surveys, such as the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST), will observe an unprecedented number density of galaxies. As a result, the majority of imaged galaxies will visually overlap, a phenomenon known as blending. Blending is expected to be a leading source of systematic error in astronomical measurements. To mitigate this systematic, we propose a new probabilistic method for detecting, deblending, and measuring the properties of galaxies, called the Bayesian Light Source Separator (BLISS). Given an astronomical survey image, BLISS uses convolutional neural networks to produce a probabilistic astronomical catalog by approximating the posterior distribution of galaxy counts and centroids. BLISS additionally includes a denoising autoencoder to reconstruct unblended galaxy profiles. We apply BLISS to simulated single-band images whose properties are representative of year-10 LSST coadds. By propagating the probabilistic detections from BLISS to its deblender, we produce per-object flux posteriors. Using these posteriors yields a substantial improvement in measured flux residuals compared to deterministic detections alone, particularly for highly blended and faint objects. These results highlight the potential of BLISS as a scalable, uncertainty-aware tool for mitigating blending-induced systematics in next-generation cosmological surveys.

Classifying young stellar objects (YSOs) in large infrared catalogs is challenging due to missing photometry and contamination from evolved stars and galaxies. We present an end-to-end classification pipeline applied to the Spitzer Extended Solar Neighborhood Archive (SESNA; Gutermuth et al. 2009), a catalog of ~8.5 million sources spanning 2MASS, Spitzer/IRAC, and Spitzer/MIPS. Missing data are addressed through SED fitting and imputation against ~2.2 million YSO models (Richardson et al. 2024), ~3000 stellar photosphere models (Castelli & Kurucz 2004), and galaxy templates. Model selection uses a Bayesian framework combining chi-squared goodness-of-fit with Gaia DR3 optical detections and a probabilistic reformulation of the Gutermuth et al. (2009) color-cut scheme. With imputed SEDs in hand, we apply a Self-Organizing UMAP (SOUMAP) — combining Self-Organizing Maps with UMAP dimensionality reduction — to cluster and classify the full SESNA catalog. The resulting map organizes sources by SED morphology in an interpretable two-dimensional representation, enabling identification of YSO subclasses, outliers, and unlabeled populations beyond the fixed color-based boundaries of traditional classification. We validate using a well-resolved Orion protostar sample from Pokhrel et al. (2023).

We present a physics-informed machine learning framework for modeling the full probability distribution of recoil velocities from black hole mergers, a key quantity that determines whether merger remnants are retained or ejected from dense stellar clusters, with direct consequences for merger rate predictions. The naive regression target is a seven-dimensional function of mass ratio and spin vectors, but exploiting a physical symmetry of the spin distributions reduces the effective input space to three dimensions, substantially improving data efficiency and out-of-distribution generalization. Because the recoil velocity exhibits a rich conditional distribution rather than a single deterministic value, we model it using a conditional normalizing flow trained on approximately 5000 numerical simulations. We further incorporate analytic structure into the feature representation to enforce correct limiting behavior.

The resulting model provides, to our knowledge, the first generative description of recoil velocity distributions valid across the full range of black hole mass ratios, and outperforms existing analytic and data-driven baselines in cross-validation. We demonstrate downstream impact: predicted black hole retention probabilities in stellar clusters shift noticeably under our learned distributions compared to previously used models.

Time-series spectra of supernovae encode detailed information about the composition and structure of stellar explosions. By combining radiative-transfer simulations with machine-learned emulators, we developed a probabilistic framework that enables Bayesian inference of ejecta properties from observed spectra.

The approach trains neural-network and Gaussian-process emulators on large grids of synthetic spectra generated with the TARDIS Monte-Carlo radiative-transfer code, reducing computational cost by orders of magnitude and enabling exploration of high-dimensional parameter spaces. 

In this talk I present the methodology and several applications across different supernova classes. 

Applied to thermonuclear Type Ia supernovae, our technique shows that events previously interpreted as distinct classes can be explained by a common explosion scenario. For stripped-envelope explosions, the framework reveals unexpected results: Type Ic supernovae, historically classified as helium-free events, show unambiguous signatures of small but measurable helium in their outer ejecta. For hydrogen-rich Type IIP supernovae, the same framework allows intrinsic luminosities to be inferred from spectral modeling, enabling independent distance measurements and new cosmological constraints.

Together these results show how AI-accelerated simulation-based inference can transform radiative-transfer modeling into a scalable framework for extracting physical insight from complex astronomical data.

As an observational science, astrophysics traditionally relies on correlations to validate theoretical models. However, correlations alone cannot determine the causal structure and direction among variables, or reveal unobserved confounding variables. Causal discovery—which infers causal relationships from purely observational data—addresses this gap. While relatively new to astrophysics, efforts to infer causal structures directly from astronomical data are now emerging.

In my talk, I present two recent applications of causal discovery in the Solar system and in galactic dynamics.

First, we address the long-standing debate on whether the colors of trans-Neptunian objects (TNOs) reflect TNO formation conditions or subsequent evolution. By applying causal discovery to two independent datasets, we establish the causal direction supporting that TNO colors are primordial. Remarkably, without prior knowledge, our model independently predicts the existence of a massive confounding body: Neptune.

Second, we employ causal discovery in galactic archaeology. Using observable stellar properties from simulations, our data-driven approach recovers a physically meaningful causal graph with latent nodes that correspond to unobservable physical quantities such as the birth radii of stars. Our causal model demonstrates excellent generalizability across various simulations and real-world observations.

The volume of galaxy imaging data continues to outpace that of the high quality spectroscopic data used to precisely measure physical parameters like redshift and specific star formation rate (sSFR). In the near future, this problem will be exacerbated as facilities like Rubin and Roman rapidly build up sky maps extending to higher redshifts and dimmer fluxes, regimes where the spectroscopic datasets provide an even more sparse and biased sampling of the galaxy population than among brighter, lower redshift objects. Image properties like photometric colors can be mapped to physical parameters using spectroscopic galaxies as training data, but in lower-dimensional color spaces there tend to be degenerate solutions, while in higher-dimensional spaces the curse of dimensionality exacerbates the deficient sampling by labeled data. However, since observed galaxy colors are highly correlated and driven by a small number of physical parameters (mainly redshift and sSFR), this is a perfect case for dimensionality reduction. I use uniform manifold approximation and projection (UMAP) to study the color–redshift relation, compressing a seven-dimensional Rubin+Roman-like color space to three dimensions and recovering a thin, densely-sampled manifold with monotonic and roughly orthogonal trends in redshift and sSFR. Position in this compressed color space maps coherently to redshift, such that it can be reliably interpolated from only a small and highly-biased subset of labeled data.