Title: Learning emergent models from $\textit{ab initio}$ many-body calculations

Abstract: We present a generalized framework, renormalized density matrix downfolding (R-DMD), to derive systematically improvable effective models from ab initio calculations. This framework moves beyond the common role of ab initio calculations as calculating the parameters of a proposed Hamiltonian. Instead, R-DMD provides the capability to decide whether a given effective Hilbert space can be identified from the ab initio data and assess the relative quality of ansatz Hamiltonians. Any method of ab initio solution can be used as a data source, and as the ab initio solutions improve, the resultant model also improves. We demonstrate the framework in an application to the downfolding of a hydrogen chain to a spin model, in which we find the interatomic separations for which a nonperturbative mapping can be made, and compute a renormalized spin model Hamiltonian that quantitatively reproduces the ab initio dynamics.