A Synthetic Image Generation and Deep Learning Framework for Entrained Droplet Detection

Abstract

Accurate prediction of two-phase annular flow behavior is essential for the safe operation of nuclear power plants and other thermal-hydraulic systems, where the entrained droplet field within the gas core directly influences mass, momentum, and energy transfer between phases. Despite decades of research, the characterization of entrained droplets continues to rely on empirical correlations developed from limited experimental datasets, and traditional droplet measurement techniques remain labor-intensive and limited in their ability to resolve the full droplet size distribution simultaneously. To address this limitation, this research develops a synthetic image generation pipeline coupled with a YOLO-based convolutional neural network for the automated detection and segmentation of entrained droplets in high-speed annular flow imagery. A physics-based droplet rendering model was developed to reproduce the characteristic backlit droplet appearance observed in experimental images, including the central caustic peak, the dark annular ring produced by total internal reflection, the gravitational shadow crescent, and the Gaussian defocus blur. Three generation methods were implemented: a constant size and blur testing pipeline, a probabilistically-sampled random generation pipeline, and a correlation-grounded Kataoka generation pipeline that uses the Kataoka droplet size distribution, the Sawant-Ishii-Mori entrainment fraction correlation, and a cosine-distributed defocus blur derived from the depth-of-field geometry of a circular pipe cross-section. A gradient contrast exclusion methodology based on the Sobel operator was developed to identify in-focus droplets for inclusion in the YOLO training labels, with the exclusion threshold calibrated against depth-of-field experimental images. YOLO model training was conducted in two campaigns. An initial campaign on a personal NVIDIA GeForce RTX 3050 Ti GPU compared the YOLOv8 and YOLOv11 nano segmentation architectures and identified an exclusion threshold of 0.6 as optimal for a 1.0 mm depth-of-field selection. An additional campaign on the Virginia Tech Advanced Research Computing Tinkercliffs cluster trained the YOLOv11 medium segmentation model on 1,000 synthetic training images and 400 validation images at 1024 x 1024 resolution, over 400 epochs and 32 batches. The cosine blur distribution with σmax = 6.0 pixels produced a YOLOv11m-seg model with bounding box precision and recall of 0.968 and 0.983 respectively, and bounding box mAP50 of 0.990 against the synthetic validation set. Synthetic-to-real droplet realism was quantified through a four-step intensity profile normalization pipeline that progressively isolates spatial scale, absolute intensity, and central-peak asymmetry contributions to the synthetic-real mismatch, and through a Laplacian of Gaussian scale-space analysis that compares the dominant defocus level present in the synthetic and real datasets without requiring per-droplet localization. The Laplacian of Gaussian calibration recovered a dominant blur level of 6.41 pixels in the real RPI dataset and 6.45 pixels in the cosine-blur synthetic dataset, which represents a σmax = 6.0 in the synthetic generation pipeline. The framework demonstrates that physically grounded synthetic image generation, when coupled with rigorous calibration against experimental imagery, produces training data of sufficient fidelity to support automated YOLO-based droplet measurement in annular two-phase flow, with direct applicability to thermal-hydraulic safety analysis in light water reactor systems.