Pulsar-Net: A Physics-Inspired Machine Learning System for Automatic Pulsar Identification

This article introduces Pulsar-Net, an astronomical machine learning project that combines feature engineering, XGBoost, threshold optimization, and SHAP interpretability analysis to automatically identify real pulsars from radio survey data. This project addresses the pain point of time-consuming and labor-intensive manual screening of candidate signals, builds a complete pipeline based on the HTRU2 dataset, provides a deployable solution, and offers an efficient tool for astronomical research.

Pulsars are important research objects in the universe, but the proportion of real pulsars among candidate signals generated by modern radio telescope surveys is extremely low (only 9.2% in the HTRU2 dataset), making manual screening inefficient and prone to omissions. The HTRU2 dataset contains 17898 candidate signals, each described by 8 statistical features (mean, standard deviation, kurtosis, skewness of the integrated pulse profile and DM-SNR curve), and class imbalance poses challenges to model training.

1. Feature Engineering: Design 11 additional features based on physical properties (signal intensity relationships, pulse profile sharpness, DM-SNR curve features, etc.), expanding to a 19-dimensional feature space, and apply signed logarithmic transformation to skewed features; 2. Model Architecture: Use logistic regression as the baseline, with XGBoost as the core (tune tree complexity, learning rate, etc., with 5-fold cross-validation); 3. Threshold Optimization: Search for the optimal threshold of 0.355 with F2 score as the target to balance recall and precision.

Test set performance: Accuracy 98%, PR-AUC 0.934, ROC-AUC 0.980, Precision 84%, Recall 90%, F1 score 0.87. SHAP analysis shows: log_kurtosis_profile (logarithm of integrated pulse profile kurtosis) is the most important feature; manually designed engineering features (e.g., pulsar_signature_score) are validated to be effective; feature interaction patterns are revealed, improving model transparency.

Provide complete deployment solutions: 1. Interactive Web Application: Built with Streamlit for online demonstration (https://pulsar-net.streamlit.app/), supporting feature upload to get real-time results and SHAP explanations; 2. Local Deployment: Python environment configuration, Jupyter Notebook to reproduce the entire workflow; 3. Model Persistence: Serialize and save the XGBoost model for direct use in batch prediction.

Pulsar-Net is a robust and interpretable automatic pulsar identification solution. Its technical highlights include domain knowledge-driven design, end-to-end pipeline, class imbalance handling, and interpretability priority. With the construction of next-generation telescopes like SKA, data volume will grow exponentially, making such tools more important. The project has clear code and complete documentation, providing a reference template for astronomical machine learning.