Mask R-CNN automates detonation cell analysis with under 2% error

A team of researchers from multiple institutions (including Mingyang Bu, Robson A. Schneider, and Deanna A. Lacoste) has developed a deep learning approach to automate the characterization of detonation-cell size distributions from soot-foil records. Traditional manual measurement is slow and subjective, while existing computer vision techniques struggle with high noise, blurred boundaries, and severe cell overlapping in real experimental images. The proposed method uses Mask R-CNN, an instance segmentation architecture, trained on a custom heterogeneous dataset that includes both numerical simulations and physical experiments, enhanced with transfer learning. This allows the model to produce accurate pixel-level mask predictions even within highly noisy flow fields.

In benchmark validations, the model shows high pixel-level agreement and strong robustness against noise. Predicted average cell sizes match manual measurements with relative errors under 2% for regular detonation conditions and under 3.5% for irregular ones. Sensitivity ablation experiments confirmed scale adaptability and led to a standardized preprocessing paradigm for image patching. Beyond extracting global average sizes, the model automatically tracks the transient spatial evolution of cell sizes along the propagation direction and quantitatively extracts high-order regularity features such as the irregularity index (RI) and standard deviation of cell deflection angles, which align with theoretical expectations. This significantly enhances the efficiency, objectivity, and detail of statistical analysis in detonation research.