Synthetic generation of realistic materials for testing of process–structure–property (PSP) relationships in computational materials science has gained significant traction over the past two decades. Generation tools continue to lag in some aspects of realism, leading to uncertainty and errors in simulating material response. The experimental collection of information to guide generation of 3D synthetic structures remains costly, time consuming, and challenging. These challenges are compounded by limitations of stereology, which permits estimation of 3D microstructural statistics from 2D observations under restrictive assumptions on constituent morphologies and size distributions and can be difficult to apply to microstructure metrics like constituent orientation distribution and clustering. This work seeks to overcome these challenges by introducing a framework for learning probable 3D microstructure statistics by minimizing the loss determined by matching statistics obtained from 2D observations via synthetic microstructure generation software (e.g. Dream.3D) and stereological principles. This framework is applied to short carbon fiber composite structures printed from an additive manufacturing process, Direct Ink Writing, with the hope that the framework could generalize to other particulate structures as well as allow for simulation and design of materials structure.