Deep Learning in Astrophysics

TL;DR

The paper surveys how deep learning can complement classical statistics in astronomy by embedding physical priors and symmetries into neural architectures, enabling learning from massive but sparsely labeled data. It surveys methodological foundations (universal approximation, architecture-induced inductive biases, and physics-informed constraints), then details cross-cutting techniques (multi-scale surrogates, simulation-based inference, anomaly detection, foundation models, RL, and LLM-based agents). It emphasizes practical benefits—data efficiency, robustness to domain shifts, and field-level inference—while cautioning about black-box critiques, the need for rigorous validation, and the risks of overclaiming super-resolution or universal transfer. The review argues for a balanced, physics-aware integration of DL with traditional methods to enhance information extraction, simulation fidelity, and autonomous scientific workflows across observational and theoretical astrophysics, while outlining concrete directions for future development and evaluation.

Abstract

Deep learning has generated diverse perspectives in astronomy, with ongoing discussions between proponents and skeptics motivating this review. We examine how neural networks complement classical statistics, extending our data analytical toolkit for modern surveys. Astronomy offers unique opportunities through encoding physical symmetries, conservation laws, and differential equations directly into architectures, creating models that generalize beyond training data. Yet challenges persist as unlabeled observations number in billions while confirmed examples with known properties remain scarce and expensive. This review demonstrates how deep learning incorporates domain knowledge through architectural design, with built-in assumptions guiding models toward physically meaningful solutions. We evaluate where these methods offer genuine advances versus claims requiring careful scrutiny. - Neural architectures overcome trade-offs between scalability, expressivity, and data efficiency by encoding physical symmetries and conservation laws into network structure, enabling learning from limited labeled data. - Simulation-based inference and anomaly detection extract information from complex, non-Gaussian distributions where analytical likelihoods fail, enabling field-level cosmological analysis and systematic discovery of rare phenomena. - Multi-scale neural modeling bridges resolution gaps in astronomical simulations, learning effective subgrid physics from expensive high-fidelity runs to enhance large-volume calculations where direct computation remains prohibitive. - Emerging paradigms-reinforcement learning for telescope operations, foundation models learning from minimal examples, and large language model agents for research automation-show promise though are still developing in astronomical applications.

Figures (12)

• Figure 1: Taxonomic hierarchy of statistical methods in astronomy. We adopt a nested framework where machine learning encompasses methods that extract patterns from data to make predictions, with deep learning (neural network-based approaches) forming a specialized subset. Top panel (red): Deep learning methods. Middle panel (blue): Classical machine learning techniques. Bottom panel (green): Statistical methods underlying both approaches. Word sizes reflect prevalence in the astronomy literature as of July 2025.
• Figure 2: Three key properties for astronomical modeling. Classical methods achieve different subsets: Gaussian Processes excel in data efficiency and expressivity but lack scalability; linear regression and Random Forests scale well and are data efficient but have limited expressivity and generalizability; basic neural networks offer scalability and expressivity but require large training datasets. Deep learning aims to achieve all three properties by augmenting neural networks with appropriate physical assumptions.
• Figure 3: Evolution of machine learning methods in astronomy. Top: 2015 taxonomy shows neural networks primarily as flexible interpolators (multilayer perceptrons, feed-forward networks) alongside classical methods. Bottom: 2025 taxonomy shows neural network approaches constitute a substantial fraction of applications, with specialized architectures (convolutional neural networks, transformers, graph neural networks) and probabilistic methods (normalizing flows, simulation-based inference, variational autoencoders) reflecting maturation beyond simple function approximation.
• Figure 4: Growth trajectories of machine learning techniques in astronomy. Top: Deep learning methods. Convolutional neural networks show adoption from 2016, followed by U-Net for segmentation (2018), autoencoders/VAEs for dimensionality reduction (2018), simulation-based inference with normalizing flows for Bayesian modeling (2020), and transformers and attention mechanism showing rapid growth after 2022. Bottom: Classical techniques showing sustained adoption. Grey shading indicates the envelope of techniques from the other category/panel for reference.
• Figure 5: Temporal evolution of statistical methods in astronomy (2000-July 2025) from astro-ph literature. Shows fraction of papers employing concepts in statistical modeling, classical machine learning, and deep learning. Papers categorized by highest-level methodology (Deep Learning $>$ Classical Machine Learning $>$ Statistical Modeling). Deep learning rises from negligible presence to $\sim$5% of papers by 2025, catching up with classical machine learning adoption.