The crystalline materials with nonlinear optical (NLO) properties are critically important for several technological applications, including nanophotonic and second harmonic generation devices. Urea is often considered to be a standard NLO material, due to the combination of non-centrosymmetric crystal packing and capacity for intramolecular charge transfer. Various approaches to crystal engineering of non-centrosymmetric molecular materials were reported in the literature. Here we propose using global lattice energy minimization to predict the crystal packing from the first principles. We developed a methodology that includes the following: (1) parameter derivation for polarizable force field AMOEBA; (2) local minimizations of crystal structures with these parameters, combined with the evolutionary algorithm for a global minimum search, implemented in program USPEX; (3) filtering out duplicate polymorphs produced; (4) reoptimization and final ranking based on density functional theory (DFT) with many-body dispersion (MBD) correction; and (5) prediction of the second-order susceptibility tensor by finite field approach. This methodology was applied to predict virtual urea polymorphs. After filtering based on packing similarity, only two distinct packing modes were predicted: one experimental and one hypothetical. DFT + MBD ranking established non-centrosymmetric crystal packing as the global minimum, in agreement with the experiment. Finite field approach was used to predict nonlinear susceptibility, and H-bonding was found to account for a 2.5-fold increase in molecular hyperpolarizability to the bulk value.