STAR-APTV: Deep learning-enabled 3D flow reconstruction in evaporating multicomponent droplets

Multicomponent droplet evaporation generates inherently three-dimensional, solutal-Marangoni flows that challenge single-camera velocimetry. We present STAR-APTV (Segmentation and Tracking Anything-based Robust Astigmatic Particle Tracking Velocimetry), a zero-shot, deep-learning-assisted, astigmatic particle tracking framework for time-resolved 3D-3C flow reconstruction with minimal optical hardware. We leverage zero-shot segmentation using SAM to detect particles in microscopic images without any task-specific labels or training. To characterize each detected particle under optical aberration, we combine shape-aware refinement using elliptic Fourier descriptors with intensity-based features within the refined mask region. We then estimate depth using an uncertainty-aware deep learning model, in which the estimated 3D trajectories are stabilized with a multi-object tracking algorithm and Kalman filter. Against a representative baseline (DefocusTracker), STAR-APTV detects up to six times more particles at high seeding density, while maintaining temporally coherent tracks, and preserving positional accuracy of particles in the presence of noise. Through synthetic validation, the proposed algorithm exhibited AEE = 0.077 px/frame and AAE = 1.45° in a known analytical flow field reconstruction. Experimental validation in two droplet regimes confirms robustness in complex, refractive samples and cross-setup transfer without any task-specific training. Among these flows, in the more challenging flow with relatively dense particle seeding, the detection rate was increased by nearly 70%, with increased retention rates and extended trajectories by almost three times compared to the conventional method. These results altogether demonstrate high-fidelity, single-camera, volumetric velocimetry in refractive, densely seeded environments, extending defocusing/astigmatic PTV toward complex droplet flows.