Do multi-step decay chains require fundamentally different feature extraction than two-body or prompt decays?

# Multi-Step Decay Chains and Feature Extraction

The short answer: Yes, multi-step decay chains do present fundamentally different challenges, but the field has developed specialized machine learning strategies to handle them.

Why Multi-Step Decays Are Different

When a particle decays in stages—say, a B meson → K* → π + K, then τ → leptons—you're no longer dealing with isolated final particles. Instead, you have a hierarchical tree structure where intermediate particles (like K* or τ) don't leave direct traces in the detector; only their decay products do [AG-2024.02-1078]. This creates a combinatorial explosion: given a pile of detected particles, how do you figure out which ones came from which step? For two-body decays, there's essentially one path. For multi-step chains, there are many plausible tree topologies to consider.

Prompt decays (particles produced directly from the collision) also differ from cascade decays (where one heavy particle decays into another). Charm hadrons, for example, can come from direct production *or* from beauty-hadron decay, and this distinction affects which kinematic features matter [AG-2025.04-1624].

How Feature Extraction Adapts

Rather than relying on simple particle properties (momentum, mass), researchers now encode the structure of the decay tree itself. One elegant approach uses a "lowest common ancestor generations" (LCAG) matrix to represent the hierarchical relationships between particles, then applies graph neural networks to learn from inter-particle correlations across multiple decay processes simultaneously [AG-2024.02-1078]. This allows the model to infer the correct tree topology.

For charm reconstruction in three-body decays like Λ_c⁺ → p + K₀_s → p + π⁺ + π⁻, machine learning models (XGBoost, DNNs) still use traditional kinematic features, but the challenge lies in disentangling the combinatorial background—many wrong pairings of the final three particles can also form a mass peak [AG-2025.04-1624].

Even in simpler rare decays, feature engineering becomes critical: B⁰ → K*⁰ τ⁺ τ⁻ requires careful selection and engineering of features to separate signal from a specific competing background, with boosted decision trees and fully connected networks both achieving >0.87 AUC [AG-2025.06-1484].

The Takeaway

Multi-step chains demand explicit representation of decay topology (not just final-state kinematics), but modern graph-based and tree-structured neural networks handle this naturally. Two-body and prompt decays can often get away with simpler feature sets, whereas cascade decays really benefit from architectures that "understand" the hierarchical structure.