Research

Foundation Model research in astronomy and cosmology accelerates scientific discovery through specialized AI. Key methods involve domain-specific fine-tuning of Large Language Models (LLMs) and developing multi-modal foundation models for complex scientific datasets, including simulations and spectroscopy. The AstroMLab series demonstrates specialized LLMs, even efficient 8B parameter models, achieve benchmark-topping performance in astronomy Q&A and reasoning, rivaling GPT-4o. InferA integrates these models as smart assistants for ensemble data. Establishing robust evaluation methodologies (e.g., EAIRA) is crucial for assessing these AI models' impact as scientific research assistants, transforming data analysis and knowledge acquisition within these domains.

Learn more about this research →

Machine learning (ML) is revolutionizing scientific discovery, particularly in astronomy and cosmology. Core methods encompass deep learning, generative adversarial networks (GANs), and probabilistic modeling. These empower tasks like accurate point spread function deconvolution, anomaly detection in astronomical images, and robust redshift estimation using synthetic spectra. ML also enables exploration of galaxy morphology, gravitational lens detection, and global field reconstruction from sparse sensor data. Crucially, research focuses on enhancing interpretability and quantifying uncertainty in models for high-energy physics and general scientific datasets. From mining literature for new cosmic associations to benchmarking AI-evolved cosmological simulations, ML accelerates scientific understanding across diverse phenomena.

Learn more about this research →

Research across Dark Matter and Cosmology employs diverse methods to probe the universe's fundamental components and structure. Weak lensing analyses, like k-cut cosmic shear with Hyper Suprime-Cam data, constrain modified gravity theories, refining our understanding of dark energy. Cosmological simulations, including multi-stream views and generative models, unveil the intricate caustic and topological structure of the dark matter web and haloes, predicting their formation drivers. Observational campaigns, from Red Clump star mapping to SPTpol's cluster surveys and future SPHEREx data, provide crucial constraints. Machine learning techniques, such as deep neural networks for kSZ peculiar velocity estimation, enhance precision in large-scale structure measurements, collectively advancing cosmological model validation and our comprehension of cosmic evolution.

Learn more about this research →

Emulation and inference research utilizes advanced surrogate models, primarily neural networks (including probabilistic NNs for uncertainty quantification) and Gaussian process emulators, to accelerate scientific discovery. These reduced-order models provide rapid, high-fidelity predictions, significantly cutting computational costs for complex simulations in cosmology and fluid dynamics. Applications include inferring cosmological subgrid parameters, predicting large-scale structure (e.g., matter power spectra for modified gravity), and modeling dynamic fluid flows. The focus is on developing differentiable, efficient, and uncertainty-aware prediction frameworks, enabling faster parameter estimation, robust inverse problem solving, and comprehensive exploration of intricate physical phenomena.

Learn more about this research →