jats:titleAbstract</jats:title>jats:pWe propose a sequential sampling approach to training statistical digital twins. This approach is relevant for real‐world engineering problems with expensive data generation. Prerequisite for building surrogates is sufficient data; however, oversampling does not improve regression accuracy. The time for data generation may be reduced by: (a) applying a classifier to improve data quality and avoid evaluation of infeasible inputs, and (b) employing dynamic sampling linked to regression quality. In dynamic sampling, the initial sampling rate is large to generate enough data for surrogate regression in a few iterations; the sampling rate gradually slows down with the improvement of the iteratively refined surrogate. A dynamic process and a steady‐state process from the field of carbon capture and utilization are used as case studies: pressure swing adsorption (PSA) and gas‐to‐liquids (GTL). The computational costs for surrogates generation are reduced by 86% for PSA and 51% for GTL, compared with employing a static sampling rate.</jats:p>