In structural health monitoring, data-driven damage identification models are typically trained and applied to individual structures, but training data corresponding to damaged states are rarely available from the specific structure under consideration. Population-based structural health monitoring aims to address this limitation by transferring knowledge between groups of similar structures; however, existing approaches lack theoretical guarantees on when this transfer of knowledge will succeed. This paper proposes a physics-informed polynomial-based flow model, motivated by eigenvalue perturbation theory, that, through a nonlinear feature transformation of the eigenvalues, explicitly separates transferable damage coefficients from structure-specific properties.The model parametrizes eigenvalue evolution under damage with a second-order polynomial in the damage severity, where the coefficients are learned from the source domain. The mathematical conditions under which the learned coefficients remain invariant within a population are derived from perturbation theory, providing a criterion for assessing transferability. Damage identification is performed as the joint estimation of location and severity through an optimization procedure. The framework is validated numerically on populations of shear building models and a two-dimensional frame, including mildly and strongly heterogeneous configurations. In all cases, it achieves high damage localization and quantification accuracy, without requiring training data from the target structure in its damage state.