Knit, with its inherent flexibility and ability to integrate bespoke material performance, creates a promising alternative to traditionally woven membranes in architectural textile applications. The CNC-knitting technology allows for the manufacturing of membranes with gradient expansion properties by numerically controlling the distribution of varied stitches. In architectural knitted structures, material programming is used to achieve complex bespoke three-dimensional surfaces at a large scale, with a minimum residual waste during continuous digital manufacturing. This permits to depart from the cut-pattern-based strategy commonly used for woven non-expandable membranes while allowing for the integration of multiple material properties in a single production process. In our research, we strategically guide the material expansion of knitted membranes in order to achieve non-developable textile surfaces by combining various stitch types informed by digital form-finding and structural analysis. As a result, membranes obtain their gradient stretch capacities under tension through the distributed material density. However, the heterogeneous irregularity of the distributed material density of CNC-knitted membranes makes it difficult to establish reliable digital simulations due to material complexity, novelty of the topic, and associated knowledge gaps. The success of simulation models relies on a thorough understanding of material properties, including their representation and translation between digital and physical environments. In particular, it is important to consider abstraction strategies to maintain computational feasibility of these models and accuracy of representation in order to reflect complex material composition. In this paper we investigate these questions through prototyping of simulation models and their calibration, in order to achieve geometrically more accurate results when designing with differentiated CNC-knitted membranes. Here we present the extension of the method for simulation and calibration of graded textiles, published earlier by the authors. The experiments are conducted on several CNC-knitted ceiling panels of varied three-dimensional geometry, where each is materially graded, and therefore stretch differently under the suspended weights. The digital simulation is calibrated towards the reduction of geometric deviation between the digital and physical artifacts with the aid of evolutionary optimization algorithms.