Research

Foundation Models are revolutionizing scientific research by specializing large language and multi-modal models for domain-specific applications. Research demonstrates adapting LLMs to interpret complex data like spectroscopy and achieve benchmark-topping performance in astronomy Q&A, even with efficient parameter counts. These domain-specialized models, exemplified by InferA and the AstroMLab series, function as intelligent assistants, streamlining analysis of cosmological ensemble and multi-modal simulation data through advanced reasoning. A critical focus is establishing rigorous methodologies for evaluating AI models as reliable scientific research assistants, ensuring their capacity for high-level contributions and accelerating scientific discovery across complex datasets.

Learn more about this research →

Machine Learning (ML) is transforming scientific research, especially in astrophysics and engineering, by addressing challenges with sparse, high-dimensional data. Key methods include neural networks, generative adversarial networks (GANs), and probabilistic modeling. These are applied to tasks such as anomaly detection, classification, and reconstruction of scientific phenomena (e.g., galaxies, cosmic structures, global fields), alongside synthetic data generation and stress field analysis. A critical focus is on enhancing model interpretability and quantifying uncertainty, vital for scientific rigor. This enables efficient analysis of vast datasets (e.g., LSST), improves predictions (e.g., redshift, cluster mass), and accelerates discovery across cosmology and other scientific domains.

Learn more about this research →

Dark Matter and Cosmology research employs diverse methodologies to probe the universe's fundamental constituents and evolution. Studies utilize multi-stream analyses of simulations to map the dark matter web's fine-grained topology and halo structures. Observational efforts, from cosmic shear analyses (Hyper Suprime-Cam) to cluster surveys (SPTpol), constrain modified gravity and cosmological parameters. Advanced techniques include modeling galaxy formation with simulations and analyzing stellar populations (e.g., Red Clump, CEMP stars via Gaia) to map galaxy evolution and structure. Machine learning (SHAMNet) provides differentiable large-scale structure predictions, enhancing parameter inference. Future missions like SPHEREx promise further insights into inflation and galaxy evolution.

Learn more about this research →

Emulation and inference research employs advanced surrogate models to efficiently analyze complex physical systems. Methods include probabilistic neural networks (PNNs) and Gaussian process emulators, often integrated with reduced-order models (ROMs) that operate in latent spaces. These techniques are crucial for accelerating simulations and enabling robust parameter inference across diverse fields. Applications span cosmological subgrid physics, modified gravity theories, and fluid dynamics, where they provide rapid predictions of high-dimensional outputs like matter power spectra or flow fields. A key impact is the ability to quantify uncertainty, handle time evolution, and enable data recovery, significantly advancing scientific discovery and design while mitigating computational costs.

Learn more about this research →