CosmicAI 2026 Seed Funding Announcement

We are pleased to announce the recipients of the 2026 CosmicAI Seed Funding Awards. The response to our call was extraordinary, with five times as many proposals as we could support. We were impressed by the overall quality, creativity, and strong research plans that thoughtfully integrated astronomy and AI.

The selected projects distinguished themselves through their originality, scientific potential, and clear alignment with our mission to advance AI-driven discovery in astronomy. We are proud to welcome these teams into the growing CosmicAI community. It is a privilege to support their work, and we look forward to the advances they will make in deepening our understanding of the universe.

• Caroline Morley, The University of Texas at Austin (Accelerated Universe)

  Title: Unlocking the Spectral Variability of Brown Dwarfs and Exoplanets with Physics-Informed Neural Networks

  Abstract: JWST’s high-resolution spectra provide unprecedented access to the physical processes in brown dwarf and exoplanet atmospheres, including their observed spectral variability driven by evolving clouds and aerosols. Retrieving atmospheric properties from these static and time-dependent spectra requires millions of line-by-line radiative transfer (RT) evaluations across hundreds of thousands of wavelengths, far beyond the practical limits of traditional Bayesian retrieval codes.

  We propose a fast, accurate physics-informed neural network (PINN) surrogate trained on outputs from a GPU-accelerated two-stream RT solver. The GPU solver supplies large line-by-line training sets, while the PINN enforces boundary conditions, flux conservation, and simplified RT residuals. To handle the high spectral dimensionality, we incorporate convolutional spectral layers that efficiently capture local line structure in a single forward pass.

  The resulting model is fully differentiable, computationally efficient, and suitable for both static and time-variable retrievals, offering a scalable path for high-resolution RT in the JWST and ARIEL eras.

• Tanmoy Laskar, University of Utah (Observable Universe)

  Title: Accelerating Radio Interferometry Error Detection with AI.

  Abstract: Next-generation interferometers (ngVLA, SKA) will produce images at a scale that makes manual artefact detection impossible. We propose a 3-step feasibility study to automate calibration/imaging error diagnosis with AI while explicitly accounting for UV sampling and weighting. Phase I starts fith real, well-characterized raw data (e.g., the CASA 3C286 tutorial set) and injects controlled faults (e.g., antenna-based phase errors, bandpass errors, RFI) into CASA.
  
  Phase II builds a labeled training set with CASA’s simtool and tclean. Phase III validates models that include: (i) a CNN for local artifacts (rings, stripes, sidelobes); (ii) a fully connected network that separates true sky structure from image-wide artifacts; and (iii) an affine-invariant GAN to handle rotations and stretching. Our deliverables are a curated, labeled UV-image dataset, an image-domain classifier that recognizes error types, and benchmarks on held-out observations. This project explores the prospect of using AI for reliable QA in future radio astronomy pipelines.

• Paul Rosen, University of Utah (Explainable Universe)

  Title: Visual Analytics for Improving Trust and Reliability of Deep Neural Networks Used in Astronomy

  Abstract: Modern astronomy increasingly relies on deep learning models to classify galaxies, reconstruct images, identify transients, analyze cosmological simulations, and generate synthetic observations, yet their internal processes remain opaque, limiting trust and reliability. We propose extending our interpretable visualization framework, ChannelExplorer, which integrates activation-level analysis, model-layer localized diagnostics, and latent-state exploration in image-based deep learning models, to evaluate deep learning models used in astronomy. We have successfully used ChannelExplorer to reveal layer-wise confusion, simulation bias, and hallucinated features in image-based networks.
  

  This project would adapt these methods to astronomical data from optical and radio surveys, cosmological simulations, and generative models. The research will provide tools to diagnose when models rely on non-physical artifacts, improperly generalize across simulators, mishandle high-dimensional spectral data, or introduce synthetic features during reconstruction. This exploratory effort directly advances CosmicAI’s Explorable and Explainable Universe scientific groups by creating trustworthy interpretability methods that prepare AI pipelines for next-generation facilities.

• Adele Plunkett & Brian Mason, NRAO (Explorable Universe)

  Title: LLM-Driven Infrastructure for Unified Astronomical Data Access and Analysis.

  Abstract: This project seeks to accelerate astronomical research by developing LLM tools that interface with cutting-edge data, supporting three students in Summer 2026. Student 1 will develop Quasar - a LLM agent for radio astronomy (NRAO/ALMA) data, increasing the scope from basic data retrieval to advanced analysis.
  
  CosmicAI funding and collaboration are essential for the computational resources needed for training and hosting open-source LLMs. Student 2 will benchmark LLM applications for multi-wavelength, multi-modal data accessibility from radio,optical and X-ray archives (NRAO/ALMA, NOIRLab, NASA), building from AstroVisBench (Joseph et al. 2025). Student 3 will apply Anthropic's Model Context Protocol (MCP) to data providers that use IVOA standards, creating an extensible framework for natural-language queries across multiple data services.
  
  The students will evaluate initial integration of the tools at NRAO and NOIRLab/Data Lab. Together, the project will provide intelligent, transparent tools for astronomers to access and analyze data using natural language queries.

• Andrew Lipnicky, NRAO (Observable Universe)

  Title: A Machine-Learning Pipeline for Automated Spectral Labeling in the ALMA Archive

  Abstract: Interferometers such as ALMA have accumulated petabytes of spectrally rich data over the past decade. While observation metadata are well characterized, fully exploiting this archive with modern AI methods requires spectroscopic featurization that does not yet exist. Many sources have been observed repeatedly with different spectral setups and spatial resolutions, but connections between these datasets are not organized holistically and therefore remain underutilized.

  We propose to combine expertise from spectroscopic analysis researchers at MIT and members of the ALMA Archive and Pipeline working groups at NRAO/ESO to develop a pipeline for automated spectral labeling of archival ALMA data. Recent machine-learning advances at MIT show promise for automated line identification in quiescent sources such as TMC-1, but have not been validated across broader use cases or integrated across multiple observations. We will build and verify a pipeline on protoplanetary disks, then investigate multi-observation enhancements and extend to additional source classes.

• Pratik Patil, The University of Texas at Austin (Explorable Universe)

  Title: Robust and Transferable Simulation-Based Inference for Explainable Cosmological Inference.

  Abstract: Modern galaxy-formation and cosmological analyses rely on expensive simulators whose likelihoods are implicit. While neural simulation-based inference (SBI) methods can be powerful, they are often difficult to interpret and can behave unpredictably under model misspecification and/or high-dimensional (especially overparameterized) parameterizations. This project will develop an interpretable SBI framework that calibrates simulators using scientifically meaningful scaling relations (e.g., Tully–Fisher, stellar–halo mass, etc.), rather than opaque learned summaries. Our batched-weighting sudo-posterior forms likelihood-free pseudo-posteriorsby kernel-weighting parameter draws based on residual discrepancies under the observed relations.
  
  In the course of the project, we will: (i) develop diagnostics and theory for misspecification and under-identification (pseudo-true sets) in overparameterized settings, and (ii) introduce a transfer-learning layer that reuses existing simulation output under prior shifts or across simulator/fidelity changes through density-ratio corrections. This project will provide an interpretable, robust, and transferable alternative to standard SBI that is tailored for astrophysics applications.

• ChangHoon Hahn, The University of Texas at Austin (Explainable Universe)

  Title: Variational Symbol Regression for a New Galaxy-Halo Phenomenology.

  Abstract: We will develop a new probabilistic phenomenology for the galaxy-halo connection that can simultaneously describe small-scale galaxy clustering predicted using different cosmology, galaxy formation, and dark matter physics. We will learn a universal functional form of the galaxy-halo connection using Symbolic Regression (SR) trained on thousands of simulations in the DREAMS suite.

  We will develop a new Variational SRmethodology that combines SR with variational inference to directly learn the functionalform of the probability distribution of galaxy position and velocity given cosmology, baryonic feedback, environment, and halo properties. The new galaxy-halo phenomenology from this work will serve as an interpretable surrogate model for hydrodynamic simulations and provide more accurate modeling of small-scale galaxy clustering.