Do neural networks need physics blueprints to spot rare Higgs decays?

Physicists built a hybrid neural network that feeds transformers both raw particle data and explicit QCD branching patterns (Lund plane) side by side. The model flags top quarks and b-jets better by exploiting structured physics, but gains nothing for simpler topologies—suggesting deep learning still benefits from human-designed physics representations. For Higgs pair hunts, the improvement is substantial enough to matter for discovery.