Scale-Invariant Jet Tagging via Relational Calculus: Zero-Shot Transfer Across Collision Energies

This paper presents a scale-invariant machine learning approach for jet tagging in High Energy Physics based on the Relational Calculus framework. Traditional models trained on absolute kinematic variables suffer from severe performance degradation when applied across different collision energies due to data drift and dependence on fixed energy scales. To address this limitation, the proposed method transforms particle kinematics into a dimensionless, Lorentz-invariant relational space by expressing constituent transverse momenta as fractions of the jet’s total transverse momentum and anchoring the system to invariant mass. This representation removes dependence on absolute energy scales and enables models to learn intrinsic decay geometry rather than environment-specific magnitudes. The approach is evaluated on the Top Quark Tagging Reference Dataset under a strict zero-shot transfer setting, where models trained on low-energy jets are tested on high-energy regimes. The relational model achieves an AUC of 0.9564, significantly outperforming a standard normalized baseline (AUC 0.8109), demonstrating strong robustness to distribution shifts. In addition to improved generalization, the framework enables high performance using lightweight models, reducing computational cost and energy consumption compared to large deep learning architectures. This work highlights the potential of relational representations for building scalable, efficient, and energy-aware machine learning systems in current and future collider experiments.