Nesar Ramachandra

Nesar Ramachandra

Computational Scientist, Argonne National Laboratory

Research

My research focuses on developing and applying computational methods at the intersection of astrophysics, cosmology, and machine learning. The work spans foundation models for scientific applications, advanced ML techniques for astronomical data analysis, cosmic structure investigation, and statistical inference methods.

Disclaimer: This section is automatically updated by Reasoning Language Models. Google Gemini is utilized to periodically go over my recent publications, talks and activities to update the content. While the information is monitored, at times incorrect information may appear.

Foundation Models

Foundation Models research in astrophysics and cosmology employs specialized AI to accelerate scientific discovery. Methods include fine-tuning large language models (LLMs) and developing multi-modal architectures to interpret data, from spectroscopy to cosmological simulations. AstroMLab initiatives demonstrate benchmark-topping astronomy Q&A performance using dedicated reasoning models (70B, 8B), matching general AI. InferA illustrates AI assistants for intricate data analysis. The impact is significant: enhancing data understanding, enabling advanced scientific reasoning, and establishing robust evaluation methodologies (EAIRA) for AI as research assistants, fundamentally transforming scientific engagement.

Learn more about this research →
Figure from Multi-modal Foundation Model for Cosmological Simulation Data
From: Multi-modal Foundation Model for Cosmological Simulation Data
7 Publications 1 Figures Available

Machine Learning for Science

Machine learning significantly impacts scientific research, notably in astronomy and cosmology. Deep learning, including neural networks and generative models, enables galaxy classification, strong lens detection, stellar characterization, and anomaly detection in vast datasets (e.g., Gaia DR3). Probabilistic modeling and interpretable uncertainty quantification enhance reliability for high-dimensional data. Multi-task learning addresses sparse engineering applications. These methods accelerate discovery, from benchmarking cosmological simulations and global field reconstruction to mining concept associations and optimizing dark energy studies, fundamentally transforming scientific inquiry across diverse domains.

Learn more about this research →
Figure from Application of probabilistic modeling and automated machine learning framework for high-dimensional stress field
From: Application of probabilistic modeling and automated machine learning framework for high-dimensional stress field
Figure from Benchmarking AI-evolved cosmological structure formation
From: Benchmarking AI-evolved cosmological structure formation
Figure from A Modular Deep Learning Pipeline for Galaxy-Scale Strong Gravitational Lens Detection and Modeling
From: A Modular Deep Learning Pipeline for Galaxy-Scale Strong Gravitational Lens Detection and Modeling
18 Publications 10 Figures Available

Dark Matter & Cosmology

Dark matter and cosmology research integrates diverse methodologies to illuminate cosmic structure and fundamental physics. Weak lensing cosmic shear analyses, exemplified by Hyper Suprime-Cam data, constrain modified gravity theories like $f(R)$, testing alternatives to $\Lambda$CDM. Advanced phase-space techniques, including multi-stream views and caustic analyses, reveal the intricate internal structures of dark matter halos and the complex topology of the cosmic web. These methods provide a detailed portrait of particle trajectories and their aggregation. Auxiliary-variable-guided generative models identify physical drivers of halo formation. Cosmological simulations, such as CRK-HACC, model galaxy formation within this dark matter framework. Furthermore, cluster surveys (e.g., SPTpol) constrain cosmological parameters, collectively enhancing our understanding of the universe's evolution and composition.

Learn more about this research →
Figure from Multi-stream portrait of the Cosmic web
From: Multi-stream portrait of the Cosmic web
Figure from Topology and geometry of the dark matter web: a multistream view
From: Topology and geometry of the dark matter web: a multistream view
Figure from The Caustic Design of the Dark Matter Web
From: The Caustic Design of the Dark Matter Web
9 Publications 4 Figures Available

Emulation & Inference

Research in emulation and inference leverages advanced machine learning to accelerate scientific discovery across diverse fields. Key methods include constructing high-fidelity surrogates and emulators—often using deep neural networks, differentiable architectures, or Gaussian processes—to replace computationally expensive simulations. These models enable rapid predictions, uncertainty quantification, and efficient parameter inference. Applications span cosmological subgrid physics, large-scale structure predictions, probabilistic redshift estimation, and peculiar velocity reconstruction, as well as reduced-order modeling for complex fluid flows. This approach significantly speeds up exploration of vast parameter spaces, facilitating robust model testing, data recovery, and scientific understanding with quantified uncertainties.

Learn more about this research →
Figure from Matter Power Spectrum Emulator for f(R) Modified Gravity Cosmologies
From: Matter Power Spectrum Emulator for f(R) Modified Gravity Cosmologies
Figure from Emulator-Based Inference of Cosmological Subgrid Models
From: Emulator-Based Inference of Cosmological Subgrid Models
Figure from Probabilistic neural networks for fluid flow surrogate modeling and data recovery
From: Probabilistic neural networks for fluid flow surrogate modeling and data recovery
8 Publications 8 Figures Available