DMAGNet: Explainable AI Empowers Astronomical Research on Galaxy Morphology Classification

DMAGNet is a convolutional neural network (CNN) designed specifically for galaxy morphology classification. Its core feature is emphasizing model interpretability while maintaining high accuracy, helping astronomers understand how AI classifies galaxy structures. This project addresses the need for AI decision-making basis in the field of astronomy, applies to scenarios such as large-scale sky survey data processing, and provides a reference for scientific AI applications.

Galaxy morphology classification is a fundamental problem in astronomy, closely related to galaxy formation and evolution. Traditional manual classification is time-consuming and cannot handle the massive data from modern sky surveys (such as SDSS, DES). AI is widely used in astronomy, but its lack of interpretability makes it difficult to meet the requirements of scientific rigor.

DMAGNet uses an optimized CNN architecture: multi-layer convolution to extract multi-scale features, spatial pyramid to capture structures of different scales, and a classification head optimized for multi-class tasks. Its core innovation is the interpretability mechanism, including attention visualization (CAM/Grad-CAM), feature importance analysis, adversarial sample analysis, hierarchical explanation, etc.

Interpretability is a basic requirement for AI in astronomy: 1. Scientific verification: Ensure AI makes judgments based on physical laws rather than noise; 2. New discoveries: Help identify features that humans have not noticed; 3. Error identification: Detect biases in training data; 4. Transferability: Facilitate understanding and adoption by other teams.

DMAGNet is applicable to: 1. Large-scale sky survey data processing, providing reliable classification labels; 2. Discovery of rare celestial objects (such as merging galaxies, gravitational lens candidates); 3. Galaxy evolution research, tracking morphological changes over time; 4. Enhancement of citizen science projects, assisting volunteers in classification.

The challenges include: data imbalance (few samples for some categories), resolution differences (cross-dataset generalization), morphological continuity (blurred category boundaries), redshift effect (blurred images of high-redshift galaxies). In the future, models need to be optimized in a targeted manner.

DMAGNet represents the direction of deep integration between AI and astronomy, emphasizing that AI needs to be "accurate" and "trustworthy". Interpretability makes AI a partner of scientists, handling massive data while showing the reasoning process. This model provides a reference for scientific AI applications and helps humans explore the universe.