Many artificial structures consist of a large number of components. To perform structural analysis of such assemblies using the finite element method (FEM), a large number of elements is required even for accurate geometric modeling, and further mesh refinement is necessary to improve analysis accuracy. Advances in parallel computing and computational performance have enabled analyses based on such large-scale, high-fidelity, and high-dimensional models (HDMs). However, further acceleration of computational speed is required to achieve more practical analyses. In recent years, reduced-order models (ROMs) have been proposed as an approach to accelerate computation by reducing the dimensionality of analysis models while minimizing the loss of accuracy, and extensive research has been conducted to apply this approach to various problems in computational mechanics. In this study, a method for detailed finite element analysis of assembly structures is proposed in which components of interest are modeled with detailed meshes (HDMs), while the remaining components are represented by Galerkin projection–based. The HDMs and ROMs are connected based on the substructuring-based domain decomposition method (primal substructuring). In the primal substructuring, the problem associated with a single subdomain is formulated as a Dirichlet boundary value problem on the subdomain. This problem is replaced with an ROM. The degrees of freedom on the interfaces between subdomains are solved using the conjugate gradient (CG) method, which is an iterative solver. The CG method exhibited good convergence for both a simple rectangular solid problem and a more complex nuclear power plant model. By this approach, the total number of degrees of freedom is reduced while the coupled interactions among multiple components are retained.