Ashkbiz Danehkar

Ash Danehkar is a Scientist in Astronomy at the Science and Technology Institute (STI) of the Universities Space Research Association (USRA). He previously held research positions at Eureka Scientific (2022–2025), the University of Michigan (2019–2021), and the Harvard–Smithsonian Center for Astrophysics (2015–2018). He received his Ph.D. in Physics and Astronomy from Macquarie University, Australia (2014), an M.Sc. in Plasma Physics from Queen's University Belfast, UK (2009), and an M.Sc. in Computational Engineering from the University of Rostock, Germany (2007). His research interests include observational astronomy, computational astrophysics, active galactic nuclei, and ionized nebulae, in addition to AI-assisted code development, deep learning, and AI agents for astronomy in NASA's Office of Data Science and Informatics (ODSI) in the Marshall Space Flight Center (MSFC) Astrophysics Branch (ST12).


Sessions

11-02
15:24
12min
xstar-tools: AI-Assisted Modernization of the XSTAR Photoionization Code
Ashkbiz Danehkar

Modernizing established astronomical software requires preserving scientific behavior while improving portability, maintainability, and integration with modern workflows. We present xstar-tools, an open-source modernization of the FORTRAN XSTAR photoionization code through an AI-assisted software development. The project progressed from a pure Python implementation to an independent C++ engine with Python interfaces and native tools for generating XSPEC spectral tables.

Large language models assisted our code development, memory and runtime optimization, testing, and documentation under human supervision. Scientific qualification used FORTRAN XSTAR as the authoritative reference and examined atomic level populations, heating and cooling rates, convergence trajectories, radiative transfer, spectra, and output products. Validation combined numerical acceptance criteria with exact comparisons against accepted baselines. Qualification diagnostics exposed subtle differences in floating-point accumulation order, atomic level indexing, and the association of physical quantities with output records. Revised implementations were accepted only after passing the relevant qualification checks, with optimization changes additionally evaluated for memory use and runtime.

The xstar-tools suite also incorporates serial and local multiprocess execution, optional MPI execution for high-performance computing, and XSPEC table generation. Native packages are qualified on Linux, macOS, and Windows, with distribution through PyPI and conda-forge on supported platforms. These capabilities facilitate deployment and integration into other scientific applications while retaining the established XSTAR physical framework.

We discuss practical lessons from AI-assisted modernization, emphasizing the role of reference implementations, human evaluation, traceable diagnostics, and reproducible acceptance criteria in producing scientifically trustworthy software. Future development of xstar-tools will extend the framework toward post-processing hydrodynamic simulations.

AI as tool for software engineering
Seminar Room 2
11-03
14:15
15min
AstroAKD: An MCP-Based Agent for Natural-Language Search of Astronomical Archives
Ashkbiz Danehkar

Astronomical data discovery often requires identifying appropriate archives, resolving astronomical object names and coordinates, translating observational constraints into service-specific query parameters, and interpreting complex metadata returned by heterogeneous data services. Although the Astroquery Python package provides programmatic access to many astronomical archives, coordinating these resources still requires detailed knowledge of individual interfaces. We present AstroAKD, a data-search agent developed within NASA IMPACT’s Accelerated Knowledge Discovery (AKD) ecosystem, a modular platform for developing, evaluating, and deploying scientific AI agents that connect natural-language reasoning with external data and software services.

AstroAKD employs a Model Context Protocol (MCP) server built on Astroquery that dynamically exposes supported archive functions to a large language model (LLM) agent. At runtime, the agent discovers available functions, inspects their parameters, converts structured requests into astronomy-specific objects such as coordinates and unit-bearing quantities, and normalizes returned tables and metadata into machine-readable responses. Supported services include SIMBAD, ADS, MAST, HEASARC, IRSA, NED, Gaia, VizieR, and Virtual Observatory resources.

Given a scientific request, AstroAKD resolves target identifiers, extracts observational constraints, selects appropriate services, constructs executable queries, and returns relevant metadata and candidate datasets with archive provenance. We describe the agent architecture and evaluate representative single- and multi-archive workflows using archive and function selection, parameter correctness, task completion, and recovery from ambiguous identifiers, empty results, and unavailable services. Benchmark requests, tool calls, and outputs are retained as inspectable traces for validation and repeatability. AstroAKD demonstrates how domain-specific AI agents and standardized tool interfaces can support transparent, rapid exploration of astronomical archives.

AI as a tool for data discovery and data management
Banquet Hall