ASTRA: An Autonomous Framework for Physics-Validated Analysis and Hypothesis Generation in Astrophysics
Abstract
We present ASTRA, an autonomous framework for physics-validated analysis and hypothesis generation in astrophysics without per-step human intervention. It integrates dimensional analysis, structural causal models, Bayesian evidence computation, and multi-wavelength data fusion in a single automated workflow. Five test cases on real observational data recover known results: Herschel filament scaling, Chandra Deep Field South cross-matching, SDSS galaxy correlations, Gaia causal inference, and Bayesian virial model selection. A sixth case runs unsupervised on synthetic galaxy data. The main sequence and mass-metallicity relation are recovered, and the outlier detector, with a false-discovery-rate threshold and no contamination parameter, recovers the blind injected fraction on ten sealed datasets (purity > 0.99) and flags nothing on the 3,000-galaxy sample. A knowledge-isolation ablation shows pattern detection and physical validation are invariant to removal of domain knowledge, which affects only interpretation. In a direct comparison, agentic-LLM and pure-ML baselines match the statistical fitting but not the physics-validation stage. They rarely derive the equilibrium benchmark, split evenly on the physical verdict, and adopt a supplied erroneous benchmark in all six trap runs, an astrophysics-specific fine-tuned model completes zero of six runs. On live SDSS DR16 data it flags 197 of 4,999 galaxies at a data-determined threshold, ranked extremes classifiable from withheld spectroscopy. These are recovery validations and capability demonstrations rather than discoveries, with hypothesis generation demonstrated qualitatively.
Related
- A semi-autonomous multi-agent pipeline for high-energy astrophysics research with enforced provenance and verifiable self-extension
- Distinguishing cosmology, astrophysics and systematics: the Bayesian search for a primordial gravitational-wave background in next-generation observatories
- DewTwin-Coin: an onboard autonomous framework for lunar water-ice prospecting using Chandrayaan-3 LIBS and ChaSTE data
- A Spectral Framework for Testing the Quasi-Star Hypothesis in Little Red Dots I: Weighing LRDs by Their Super-Eddington Luminosity Ratios---No Signs of Overmassive Black Holes
- A Unified AI Framework for Rapid and Accurate Waveform Generation of Compact Binary Coalescences
BibTeX
@misc{white2026astra,
title = {{ASTRA: An Autonomous Framework for Physics-Validated Analysis and Hypothesis Generation in Astrophysics}},
author = {Glenn J. White},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.36144},
eprint = {2610.36144},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.36144},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - ASTRA: An Autonomous Framework for Physics-Validated Analysis and Hypothesis Generation in Astrophysics
AU - Glenn J. White
PY - 2026
DA - 2026/10/26
PB - arXiv
DO - 10.48550/arxiv.2610.36144
UR - https://arxiv.org/abs/2610.36144
AB - We present ASTRA, an autonomous framework for physics-validated analysis and hypothesis generation in astrophysics without per-step human intervention. It integrates dimensional analysis, structural causal models, Bayesian evidence computation, and multi-wavelength data fusion in a single automated workflow. Five test cases on real observational data recover known results: Herschel filament scaling, Chandra Deep Field South cross-matching, SDSS galaxy correlations, Gaia causal inference, and Bayesian virial model selection. A sixth case runs unsupervised on synthetic galaxy data. The main sequence and mass-metallicity relation are recovered, and the outlier detector, with a false-discovery-rate threshold and no contamination parameter, recovers the blind injected fraction on ten sealed datasets (purity > 0.99) and flags nothing on the 3,000-galaxy sample. A knowledge-isolation ablation shows pattern detection and physical validation are invariant to removal of domain knowledge, which affects only interpretation. In a direct comparison, agentic-LLM and pure-ML baselines match the statistical fitting but not the physics-validation stage. They rarely derive the equilibrium benchmark, split evenly on the physical verdict, and adopt a supplied erroneous benchmark in all six trap runs, an astrophysics-specific fine-tuned model completes zero of six runs. On live SDSS DR16 data it flags 197 of 4,999 galaxies at a data-determined threshold, ranked extremes classifiable from withheld spectroscopy. These are recovery validations and capability demonstrations rather than discoveries, with hypothesis generation demonstrated qualitatively.
ER -