Research

Foundation Model research in astronomy centers on developing specialized AI for scientific inquiry. Key methods involve creating multi-modal foundation models for complex cosmological simulation data and adapting Large Language Models (LLMs) to grasp domain-specific knowledge, including spectroscopy. Projects like AstroMLab demonstrate achieving benchmark-topping and GPT-4o level performance in astronomy Q&A, utilizing domain-specialized reasoning models ranging from 8B to 70B parameters. This research establishes rigorous methodologies for evaluating AI as scientific research assistants. The impact is profound: these models, exemplified by InferA, function as intelligent assistants, significantly enhancing data analysis, accelerating discovery, and bridging the gap between advanced AI capabilities and specialized scientific needs.

Learn more about this research →

Machine learning is revolutionizing scientific research, particularly in astronomy and cosmology, by analyzing vast datasets from surveys like Rubin LSST, Gaia, and Hyper Suprime-Cam. Key methods include deep learning (neural networks, GANs), probabilistic modeling, and unsupervised techniques. Applications span anomaly detection in astronomical images, robust galaxy selection for weak lensing cluster mass estimation, and strong gravitational lens identification. ML also enables accurate redshift and peculiar velocity estimation, enhances cosmological structure formation simulations through physical benchmarking, and reconstructs global fields from sparse sensors. A strong focus is placed on enhancing interpretability, quantifying uncertainty in AI models, and leveraging multi-task learning to improve scientific discovery and physical inference.

Learn more about this research →

Recent research in Dark Matter & Cosmology employs diverse methods to elucidate cosmic structures. Multi-stream and phase-space analyses unveil the intricate internal dynamics, topology, and caustic features within dark matter haloes and the cosmic web. Advanced cosmological simulations, such as CRK-HACC, model the co-evolution of dark matter and baryonic matter, linking distribution to galaxy formation. Machine learning, specifically generative models, uncovers physical drivers of halo structure. Concurrently, cosmic shear analyses, using data from Hyper Suprime-Cam, constrain alternative gravity theories like f(R) gravity, testing fundamental cosmological paradigms and their impact on large-scale structure formation.

Learn more about this research →

Emulation and inference research develops efficient surrogate models for computationally intensive physical simulations. Key methods include neural networks, often designed to be probabilistic for robust uncertainty quantification, and Gaussian processes. These are applied across diverse domains such as cosmology, for inferring subgrid model parameters and predicting matter power spectra, and fluid dynamics, through reduced-order surrogates. Differentiable emulators further enable efficient gradient-based optimization and parameter exploration. The impact is profound: transforming complex, high-fidelity simulations into fast, accessible tools. This accelerates scientific discovery, facilitates comprehensive parameter inference with quantified uncertainties, and empowers deeper understanding in fields reliant on large-scale computational modeling.

Learn more about this research →