Learning Particle Physics Patterns Without Simulation

Summary & Methodology Analysis

ShellFlow implements a Riemannian conditional flow matching (RCFM) objective to model particle kinematics. Unlike traditional methods that rely on Monte Carlo event generators, this architecture treats particle geometry as a product of per-particle mass shells. It employs a dual-head transformer as its backbone, utilizing specialized coordinate charts for particle properties: a direction chart on S^2 for momentum direction, alongside log-energy and log-mass charts. Physics priors, such as the invariant-mass formula and the on-shell condition, are hard-coded into the model geometry and the loss function, which includes an auxiliary K-body loss term for particle groupings (K=2, 3, 4). A distribution-matching weight (w*) is applied to optimize gradient signaling, specifically addressing areas where the generator under-produces or overshoots the data distributions. The model is conditioned on event-specific features like particle type, charge, and MET, allowing it to generate kinematics consistent with the supplied composition. Current limitations include a lack of self-contained MET sampling, requiring downstream reliance on provided conditioning, and incomplete learning of joint quality-cut acceptance for specific peaks like the Higgs or Z->4l. It remains to be determined if the model captures fundamental physics or simply replicates correlations within the training statistics, and the current log-energy chart normalization reduces effective resolution for light mesons by compressing the sub-GeV regime.