Intra-source diversity in FRB 20220912A: An energetic trichotomy from hybrid machine-learning

Fast radio bursts (FRBs) are millisecond-duration cosmic transients of unclear physical origin. To probe the emission physics of a repeating source, we performed a hybrid machine-learning analysis of 1061 bursts from the hyperactive repeater FRB 20220912A---a source with a clean local environment by FAST. We applied dual-path nonlinear dimensionality reduction using uniform manifold approximation and projection and t-distributed stochastic neighbor embedding to six parameters (width, waiting time, peak frequency, bandwidth, flux, and energy), followed by density-based consensus clustering to identify potential burst subgroups without predefined labels. Mutual information and Shapley Additive exPlanations (SHAP) quantified feature importance. This approach identified a trichotomy: a low-energy cluster (CC1, 852 bursts), a moderate-energy cluster (CC2, 135 bursts), and a high-energy cluster (CC3, 71 bursts), with three bursts showing inconsistent classifications. Globally, energetic parameters dominate the classification, while SHAP reveals distinct parameter hierarchies across clusters. For CC1 and CC3, width shows SHAP trends opposite to those of energy-based parameters. For CC2, width is bidirectional. Temporal analysis shows distinct activity patterns. CC1 persisted throughout with nonstationary burst rates and clustered waiting times; CC2 was last detected at ∼MJD 59912 with nonstationary bandwidth and lengthening waiting times; CC3 exhibited episodic activity, last detected at ∼MJD 59918 with random waiting times within episodes. All clusters exhibit stochastic energetic parameters. This energetic trichotomy, observed in a clean environment, indicates multiple emission regimes within a single engine, with CC2 bridging the low-energy and high-energy extremes. Our study establishes a robust, interpretable framework for identifying subclasses of bursts in FRBs and other repeating transients.