Neural Modal Decomposition: Architectural Priors from Observables

This paper presents Neural Modal Decomposition (NMD), a method that leverages architectural priors from physical observables to decompose multi-port linear time-invariant system response matrices into pole-residue form. The approach addresses modeling challenges for diverse physical systems including RF cavities, photonic devices, and superconducting quantum chips, which share common mathematical structures despite their different underlying physics. NMD enables accurate prediction of system responses for arbitrary geometries and port configurations by learning intrinsic resonant modes from data.