This paper proposes a systematic approach to create digital twins of infrastructure for regional risk and resilience analysis. A digital twin consists of a virtual representation of infrastructure intended for specific analyses (e.g., in reliability and resilience analyses considering relevant hazards.) We formulate creating digital twins as a model selection problem whose objective is to ensure predicting the response quantities of interest (e.g., infrastructure performance or resilience measures) with the desired accuracy level under computational resources constraints. The virtual representation requires collecting and integrating data about infrastructure physical and operational characteristics from multiple sources. The required data depend on the considered analyses to predict the response quantities of interest (e.g., considering only reliability analysis or also including functionality analysis.) The collected data are typically unstructured and incomplete; so, they need to be processed and synthetically augmented to, for example, capture infrastructure’s future developments. Creating digital twins also entails deciding the scales, boundaries, and resolution of the virtual representation and selecting models for intended analyses from multiple candidates, each of different computational fidelities and evaluation costs. A digital twin is hardly a perfect representation of infrastructure reality; there are missing or limited data about infrastructure, and several sources of uncertainty affect predicting infrastructure’s states for decades ahead. Uncertainty propagation is an integral part of creating digital twins to understand how missing data and different sources of uncertainty affect predicting the response quantities of interest. Uncertainty propagation requires evaluating the created digital twins at multiple realizations of the sources of uncertainty. However, using a detailed digital twin with high-resolution and high-fidelity models can lead to a high computational cost. The proposed approach guides the creation of statistically equivalent digital twins following two general principles: 1) the selected scale, resolution, and computational fidelities collectively ensure the desired accuracy in predicting the response quantities of interest, and 2) the allocation of computational resources is based on each contribution to the uncertainty of the response quantities of interest.