Unveiling Hidden Clustering: An Unsupervised Machine Learning Study of Repeating FRB 20220912A

TL;DR

This study tackles the challenge of classifying repeating FRBs by applying an unsupervised learning pipeline that combines UMAP for dimensionality reduction and HDBSCAN for density-based clustering on eight spectral and timing features from FRB 20220912A. The analysis identifies three physically distinct clusters, including a broadband/high-fluence group that mirrors non-repeating FRB properties, and finds cross-source similarities with FRB 20201124A and FRB 121102, suggesting possible shared emission mechanisms. Robustness checks—including waiting-time removal, sensitivity analyses, and alternative embedding methods—support the three-cluster taxonomy and reduce the likelihood that results are cadence-driven. The results imply that some non-repeating FRBs could be bright repeats obscured by observational limitations, and they favor curvature radiation scenarios with emission-height–dependent frequencies over simple maser models. Overall, the work demonstrates the value of unsupervised clustering in uncovering intrinsic FRB diversity and informs future classification efforts and physical modeling.

Abstract

Fast Radio Bursts (FRBs) are millisecond-duration radio transients of extragalactic origin. Classifying repeating FRBs is essential for understanding their emission mechanisms, but remains challenging due to their short durations, high variability, and increasing data volume. Traditional methods often rely on subjective criteria and struggle with high-dimensional data. In this study, we apply an unsupervised machine learning framework that combines Uniform Manifold Approximation and Projection (UMAP) and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) to eight observed parameters from FRB 20220912A. Our analysis reveals three distinct clusters of bursts with varying spectral and fluence properties. Comparisons with clustering studies on other repeaters show that some of our clusters share similar features with sources such as FRB 20201124A and FRB 121102, suggesting possible common emission mechanisms. We also provide qualitative interpretations for each cluster, highlighting the spectral diversity within a single source. Notably, one cluster shows broadband emission and high fluence, which are typically seen in non-repeating FRBs. This raises the possibility that some non-repeaters may be misclassified repeaters due to observational limitations. Our results demonstrate the utility of machine learning in uncovering intrinsic diversity in FRB emission and provide a foundation for future classification studies.

Figures (9)

• Figure 1: The Silhouette Score and Davies-Bouldin Score for different values of n_neighbors. For clustering, a higher Silhouette Score and a lower Davies-Bouldin Score indicate better performance. As shown in this figure, when n_neighbors = 6, the Silhouette Score reaches its maximum and the Davies-Bouldin Score reaches its minimum. Therefore, we choose n_neighbors = 6 for the following analysis.
• Figure 2: (a) Silhouette Scores as a function of t-SNE perplexity, maximized at perplexity = 19. (b) Silhouette Scores as a function of t-SNE early_exaggeration, maximized at early_exaggeration = 12. (c) Silhouette Scores as a function of KMeans n_clusters after t-SNE embedding, maximized at n_clusters = 3. (d) Silhouette Scores as a function of KMeans n_clusters after PCA embedding, maximized at n_clusters = 2.
• Figure 3: The unsupervised machine learning analysis of FRB 20220912A also yields three clusters, consistent with previous studies.
• Figure 4: Parameter coloring of the clustering result for n_neighbors = 6.
• Figure 5: The histograms (a)–(c) show the distributions of cluster count, noise fraction, and ARI, along with their mean values obtained from 30 random seeds under the base hyperparameter settings. The heatmaps (d)–(f) show the UMAP sensitivity check results for n_neighbors ranging from 4 to 8 and min_dist ranging from 0 to 0.1, including cluster count, noise fraction, and ARI scores relative to the base results. The heatmaps (g)–(i) show the HDBSCAN sensitivity check results for min_cluster_size ranging from 80 to 120 and min_samples ranging from 8 to 12, including cluster count, noise fraction, and ARI scores relative to the base results.