Beyond Classification and Prediction - The Promise of Physics-Informed Machine Learning in Astronomy and Cosmology

Aug 3, 2026

Uploaded by Aurora

Machine learning in astronomy and cosmology faces a central challenge: many remain concerned that standard algorithms are opaque and therefore limit scientific usefulness. Physics-informed machine learning (PIML) addresses opacity by embedding physical laws and domain knowledge during training and/or by shaping neural network architectures around those principles. This chapter examines two representative PIML uses, highlighting gains in predictive accuracy and efficiency, while arguing that interpretability improves only in the second approach.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages20

Reading Time50 min

AccessFree to download

Introduction

Physics-informed machine learning as a next-generation approach

Incorporating physical laws and domain knowledge

Applications in astronomy and cosmology

Machine learning in Astronomy and Cosmology

Beyond Classification and Prediction: The Promise of Physics-Informed Machine Learning in Astronomy and Cosmology Helen Meskhidze∗ Written for: Juan Dur´an and Giorgia Pozzi (Eds.): Philosophy of Science for Machine Learning: Core Issues and New Perspective, Synthese Library. Abstract Though the use of machine learning (ML) is ubiquitous in astrophysics and cosmology, many still see the opacity of ML algorithms as a major issue to their scientific utility. One way of addressing this opacity is through an emerging trend in ML research of “teaching” ML algorithms physical laws and domain-specific knowledge. “Physics-informed machine learning” (PIML), as this methodology is called, promises to produce better predictions and yield more interpretable algorithms. It does so by using physical principles in the training process and/or by using physical principles to guide the development of the neural network architecture. In this chapter, I investigate two uses of PIML in astronomy/cosmology, each a representative example of the two PIML methods. In both cases, PIML provides improvements in terms of the predictions and efficiency of ML algorithms. However, I argue that only in the second case does PIML offer any improvement in terms of the interpretability of the algorithms. Keywords: opacity, interpretability, understanding, machine learning, neural network, astronomy, cosmology 1 Introduction “Physics-informed machine learning” (PIML) has been called “the next generation of artificial intelligence” by popular media (Andrzejczuk 2023) and the scientists who use it claim that it can “transform our modeling, simulation, and understanding of complex physical systems in various science and engineering disciplines”(Chen, Liu, and Sun 2021) . PIML aims to incorporate physical laws and domain-specific knowledge into machine learning in order to produce better predictions and yield more interpretable algorithms. Domain-specific knowledge, in this context, can ∗ Black Hole Initiative, Harvard University, 20 Garden Street, Cambridge, MA 02138, [emeskhidze@g.harvard.edu](emeskhidze@g.harvard.edu) include symmetry laws, conservation laws, or even the governing dynamics of a system. Two means have been proposed for the incorporation of such physical laws and domain-specific knowledge: by “teaching” machine learning algorithms this information and/or by designing specialized network architectures. In either case, physics-informed machine learning is argued to yield better predictions in the presence of imperfect or noisy data because the algorithms are more robust to any small irregularities in the data (Karniadakis, Kevrekidis, Lu, et al. 2021, 423) . More importantly, though, PIML is thought to be more interpretable because our knowledge of the physical world is incorporated from the start. Since their introduction in Raissi, Perdikaris, and Karniadakis (2019), PIML methods have been developed and used in numerous domains across the sciences. Some examples of projects include performing parameter estimation for applications in systems biology (Daneker, Zhang, Karniadakis, and Lu 2023), modeling fluid dynamics (Cai, Wang, Fuest, Jeon, Gray, and Karniadakis 2021), characterizing a crack in the surface of a material (Shukla, Di Leoni, Blackshire, Sparkman, and Karniadakis 2020), predicting the many-electron wave equation for applications in quantum chemistry (Pfau, Spencer, Matthews, and Foulkes 2020), and forecasting weather/climate processes (Kashinath, Mustafa, Albert, et al. 2021) . In astronomy and cosmology, PIML has been used to, e.g., model the formation of molecular clouds in the interstellar medium (Branca and Pallottini 2023), solve the radiative transfer equation for supernova simulations (Chen, Jeffery, Zhong, et al. 2022; Mishra and Molinaro 2021), investigate astrophysical shocks (Moschou, Hicks, Parekh, et al. 2023), and model the gravitational fields around small astrophysical objects (Martin and Schaub 2022a

FAQs

What problem does physics-informed machine learning aim to solve in astrophysics and cosmology?

It addresses concerns about the opacity of standard ML algorithms by incorporating physical principles and domain knowledge into the learning process, improving scientific utility.

How does PIML implement physical knowledge?

It can “teach” physical laws to the algorithm during training and/or guide neural network architecture by building in physical principles.

Do both PIML case studies improve interpretability?

No. The chapter argues that interpretability improves only in the second case study, not when physical principles are included solely through training.

Aug 3, 2026

Uploaded by Aurora