Automated Detection and Characterization of Aluminum Agglomerates in Solid Propellant Combustion

Aluminum agglomeration in solid rocket propellants has a significant impact on two-phase flow losses and combustion efficiency. Characterizing these agglomerates typically relies on high-speed photography, where manual analysis is labor-intensive and prone to subjectivity, while traditional image processing algorithms struggle with noise and motion blur. This study presents an automated pipeline for the detection of agglomerates using a Mask R-CNN deep learning architecture with a ResNet-50 backbone. The model was trained and evaluated on a dataset of 166 high-speed images containing over 8100 manually annotated instances, captured under varying pressure conditions (2–4 MPa). A shortest-chord diameter estimator is applied uniformly to all segmentation masks to reduce the systematic inflation that area- and perimeter-based estimators exhibit on motion-blurred masks. Detection performance reaches an [Formula: see text]-score of 0.82, and population-level agreement with the manual baseline yields a 1% difference in terms of [Formula: see text] (Sauter mean diameter) and a 3% difference in terms of [Formula: see text] (volume-weighted diameter). Our results demonstrate strong alignment with manual annotations, establishing deep learning as a highly scalable and effective alternative to manual analysis in this imaging regime.