International Astronomy and Astrophysics Research Journal

The rapid expansion of large-scale astronomical surveys, multi-messenger observations, and high-resolution numerical simulations has transformed astronomy and astrophysics into increasingly data-intensive sciences, driving the widespread adoption of artificial intelligence (AI) and machine learning (ML) for scientific analysis. Machine learning has achieved notable success in applications including stellar parameter estimation, galaxy morphology classification, exoplanet detection, gravitational-wave analysis, and cosmological inference. However, conventional data-driven approaches often rely on statistical correlations, limiting their physical interpretability, robustness under distribution shifts, and ability to generalise to sparsely sampled or previously unseen astrophysical conditions. These challenges have stimulated growing interest in physics-informed machine learning (PIML), which integrates physical laws, governing equations, conservation principles, and simulation-based knowledge into the learning process to improve scientific consistency and predictive reliability. This methodology-oriented critical review examines recent developments in PIML across major areas of astronomy and astrophysics, including stellar evolution, galaxy formation, cosmology, gravitational-wave astronomy, solar and space weather forecasting, and astrophysical fluid dynamics. Rather than providing an application-based survey, the review introduces a unified classification framework that organises existing methodologies according to their degree of physics integration, encompassing physics-guided learning, physics-constrained learning, physics-informed neural networks, neural operators, hybrid scientific machine learning, and emerging digital twin concepts. Using this framework, the literature is critically evaluated in terms of predictive capability, physical consistency, interpretability, computational scalability, uncertainty quantification, and methodological maturity. The review identifies several challenges that continue to limit broader adoption, including the scarcity of high-quality observational data, the computational cost of physics-constrained optimisation, the absence of standardised benchmarking practices, and limited validation under out-of-distribution astrophysical conditions. It also discusses emerging research directions, such as neural operators, multi-fidelity learning, uncertainty-aware AI, scientific foundation models, and digital twins, while emphasising that many of these technologies remain at an early stage of development for astronomical applications and require rigorous validation. Overall, the analysis indicates that increasing the integration of physical knowledge generally enhances scientific reliability, interpretability, and extrapolation capability, although these improvements are accompanied by greater computational complexity and methodological challenges. By providing a structured, methodology-oriented synthesis of the current literature, this review offers a critical perspective on the evolution of physics-informed AI and outlines key directions for developing scientifically reliable, interpretable, and physically grounded machine learning frameworks for next-generation astronomy and astrophysics.