CosmicAI 2025 Seed Funding Announcement
We’re excited to announce the recipients of our CosmicAI Seed Funding Awards for 2025. We were overwhelmed by the response to our call and received four times more proposals than we could fund, and overall were impressed by the quality, creativity, and strong research plans that combined astronomy and AI.
The selected proposals stood out for their creativity, scientific promise, and alignment with our mission to advance AI-driven discovery in astronomy. We are proud to welcome them to the growing CosmicAI community. We’re honored to support their work and look forward to seeing them take the next leap forward in furthering our understanding of the cosmos.
-
Title: Provably-Accurate, Structure-Preserving Operator Learning for Astrophysical Systems
Abstract: Operator learning for surrogate modeling offers a promising avenue to simulate complex astrophysical phenomena. We propose a scalable, provably-accurate framework that employs structure‐preserving operator learning through combinations of analytically-designed kernels judiciously weighted by neural networks. In our formulation, a surrogate operator is constructed as trainable shifts of kernels that each satisfy conservation laws, analytically-enforced boundary conditions, divergence‐free or curl‐free properties, and established boundary layer relationships. These kernels inherently support multiscale features and capture turbulence energy cascade conditions. Our proposed method naturally forms the core of an operator-valued Gaussian process, enabling both generative modeling and uncertainty quantification. This integrated approach will deliver computational efficiency while rigorously preserving physical laws, advancing the fidelity and interpretability of astronomical simulations. By bridging data‐driven techniques with principled physics constraints, our proposed approach will address critical challenges in modeling nonlinear, high‐dimensional astrophysical systems and promises to transform predictive accuracy in astronomical research and drive scientific discovery
-
Title: AI for accelerated simulations for an Explainable Universe
Abstract: This collaboration between astronomer Paul Torrey and AI computer and data scientist Judy Fox aims to advance our understanding of dark matter, galaxy formation, and the overall structure of the Universe. In the first phase of our research, we will employ state-of-the-art methods to create enhanced AI surrogates that are trained on both current and newsimulations (as from the DREAMS project) and enable orders of magnitude faster exploration of new hypotheses. We specifically plan to research Latent Diffusion Models, Variational Autoencoders, and Physics-Informed Neural Networks. In the second phase, these enhanced surrogate simulations will be analyzed using large-scale graph neural network structures to describe merger trees. This enables us to map the Universe and study the nonlinear relationships between global properties and galaxy dynamics. The time dependence in the merger history will allow us to relax the equilibrium assumption in causal inference in collaborative research with the Explainable Universe group and Arya Farahi from Texas
-
Title: From Rags to Riches: Analyzing Merger Trees to Predict Rich Simulation Results from Simple Ones
Abstract: Rich, fluid dynamics-based (Hydro) simulations enable astronomers to study the effects of different physics on the formation of galaxies and other celestial bodies, but are costly to run. Simple, Dark Matter Only(DMO) simulations are much cheaper, but yield limited information. We propose to utilize graph-based AI tools to predict the results of Hydro simulations from their DMO counterparts, by making use of the galaxies’ merger trees. Merger trees encode galaxy structure by recording the formation of galaxies as a sequence of discrete merging events, and contain a wealth of information about the galaxies. However, these merger trees are complex data structures and very different from Euclidean data, necessitating the use of AI tools. This approach promises to significantly reduce the cost of new simulations, while improving our understanding of the relationship between galaxies’ formation history and their physical properties
-
Title: Effective Confidence Estimation for Long-Form Generation with Large Language Models (LLMs)
Abstract: Confidence Estimation (and relatedly, performance prediction and uncertainty quantification) predicts the quality of model outputs. This enables calibrated trust, letting human users know how much to trust model outputs for use, oversight, and human-AI teaming. Confidence estimation has been traditionally framed as predicting the probability that a model output is correct or not. This binary framing is reasonable for traditional classification tasks but overly simplistic for long-form LLM generation, wherein fine-grained, partial-credit evaluation is more appropriate. For example, predicting that LLM-generated programming code is 95% correct (needing minimal human correction) would be far more useful than predicting confidence that generated code is 100% correct or not. This research project will develop, across LLMs and long-form generation tasks (e.g., code generation), the capability to predict task-specific performance metrics. The ultimate goal is to accelerate discovery in astronomy and other scientific workflows by enabling more effective LLM use and human-AI teaming.
-
Title: Developing Training LLM Notebooks for NRAO Archival Data
Abstract: We will hire 2 UVA undergraduate students who will work with UVA faculty and NRAO staff to develop well-annotated Jupyter notebooks dealing with NRAO archival data. Initial tasks will be to adapt introductory/tutorial notebooks to ensure they have appropriate labels. As time permits, the students will be asked to develop increasingly complex radio data access, reduction, and analysis tasks. Ideally, as time permits, the students will create working notebooks capable of re-creating the full body of results from recently published papers (with those published by UVA/NRAO team members being easy initial targets). As helpful, the students will connect with CosmicAI LLM experts both at NOIRLab and UT Austin to ensure the developed notebooks are in a form that would be useful for potential future use with co-pilot development. We envision this as a summer program, but would work with the Explainable Universe working group to discuss whether a sustained academic year element would be beneficial.
-
Title: A Foundation Model for Trustworthy Astronomical Source Classification with Self-Supervised Learning
Abstract: Astronomical datasets have expanded dramatically in size and complexity over the past two decades, offering vast opportunities for exploration but posing challenges due to limited labeled data. We propose leveraging self-supervised learning frameworks, specifically VICReg and JEPA, to build a foundation model that jointly extracts information from imaging and spectroscopic data. This multimodal approach, trained on the 100TB PolymathicAI database (augmented with 20 million new DESI spectra) spanning five modalities, will capture rich, shared representations of astronomical objects. Our prototype already performs preliminary classifications based on a clustering analysis. For trustworthiness, we will develop scientific metrics assessing accuracy, robustness, and interpretability. Following the assessment, we will design downstream task-specific models for refined object classification and anomaly detection to demonstrate the capability of our foundation model. This model will serve as a versatile starting point for other downstream tasks and enable a wide range of generative AI applications beyond this seed project.
-
Title: AlphaCal: An autonomous agent for processing interferometric calibrator-source datasets
Abstract: Data from an interferometer requires significant processing before astronomers can extract scientific insights. The selection and sequencing of calibration and analysis actions is a complex task, dependent on the expertise of astronomers who consider factors such as data characteristics, instrument knowledge, computational cost, and best practices. We apply reinforcement learning (RL) to this task, allowing an agent to autonomously explore and identify optimal decisions based on an objective function with metrics that quantify best practices. The strategies learned through RL are generalized and applicable for data-driven processing. With this seed funding, we will apply our proof-of-concept to data from the VLA calibrator survey. This target selection minimizes algorithmic complexity while bridging the gap between simulation and real data. A successful outcome will be a model for data-driven calibration that can be integrated into future VLA calibrator monitoring surveys.
-
Title: Efficient Event Detection in Hyperspectral Astronomy via Transfer Learning from Video
Abstract: Hyperspectral astronomical data and video data share a similar high-dimensional structure, both forming spatiotemporal or spectral cubes. While video analysis benefits from abundant annotations and well-optimized deep learning models, hyperspectral event detection struggles with data sparsity and limited labels. This project proposes adapting pre-trained video detection models for hyperspectral astronomical event detection through transfer learning, leveraging self-supervised feature extraction methods and extensive video datasets to mitigate annotation challenges. By optimizing training and inference efficiency, we aim to enhance detection performance while reducing the need for large labeled astronomical datasets. The research will explore domain-specific fine-tuning and architectural modifications to bridge the gap between video and hyperspectral feature representations. Key challenges include handling sparsity, ensuring model adaptability, and maintaining computational efficiency. This approach promises to accelerate the adoption of machine learning in astronomy, improving automation and facilitating new discoveries in hyperspectral event detection.