Deep Learning-Based Tracking and Lineage Reconstruction of Ligament Breakup
A deep learning pipeline tracks how liquid jets shatter into droplets — with implications for fuel sprays, pharmaceuticals, and fire suppression.
The Thesis
When a liquid jet hits a surface or another jet, it spreads into a thin sheet that tears apart into ligaments and droplets — a process called atomization. Understanding exactly how that breakup happens, and in what sequence, matters for designing better fuel injectors, drug inhalers, and agricultural sprayers. Conventional tracking software assumes each object becomes exactly one object in the next frame, which breaks down entirely when a single ligament shatters into five droplets at once. This paper builds a two-stage system: first, a neural network detects and classifies objects in high-speed camera footage; second, a separate model explicitly maps one-to-many 'parent-child' fragmentation events across frames. The catch is that the system was trained and tested on a single fluid (Carbopol gel) under one set of nozzle conditions, so generalization to other spray regimes remains unproven.
Catalyst
High-speed shadowgraphy cameras (cameras that capture silhouettes using backlighting) now produce frame rates and resolutions that make automated frame-by-frame analysis tractable but overwhelming for human analysts. At the same time, transformer-based sequence models — originally developed for natural language — have become practical enough to apply to short temporal windows of physical events, enabling the association logic this problem requires.
What's New
Prior multi-object tracking (MOT) frameworks such as SORT and DeepSORT maintain strict one-to-one associations between frames: each detected object maps to exactly one object in the next frame. That assumption is fundamentally incompatible with fragmentation, where a single parent ligament produces multiple child droplets simultaneously. This paper introduces an explicit three-class association model — continuation, fragmentation, and non-association — and uses physics-informed geometric features (shape, velocity, proximity) rather than raw pixels to make those assignments, achieving perfect recall on fragmentation events in its test set.
The Counter
The entire pipeline was validated on one fluid (Carbopol gel, a thickened polymer used as a model fluid) under one impinging-jet configuration — which is a very narrow experimental envelope. Real spray systems involve fuels, water, emulsions, and air-blast atomizers with vastly different viscosities, surface tensions, and breakup regimes; there is no evidence the model transfers to any of them. The synthetic data strategy augments training images without introducing 'physically implausible' configurations, but the paper does not rigorously test whether those synthetic images actually improve performance on out-of-distribution real fluids. Perfect recall (1.00) on fragmentation events sounds impressive, but with severe class imbalance and a small held-out test set, a single threshold tweak could have achieved that at the cost of massive false positives. Finally, integrating this into an actual injector development workflow requires not just detection accuracy but real-time throughput and interoperability with existing CFD (computational fluid dynamics) pipelines — none of which is demonstrated here.
Longs
- GE Aerospace (GE) — fuel atomization affects combustor efficiency in jet engines
- Nordson Corporation (NDSN) — precision spray and fluid dispensing equipment
- CECO Environmental (CECO) — emissions control tied to combustion and spray processes
- Watts Water Technologies (WTS) — spray nozzle and fluid control systems
Shorts
- Vendors of manual or semi-automated spray analysis software (e.g., LaVision SprayScan) — automated lineage reconstruction reduces reliance on expert-guided annotation workflows
- Consultancies specializing in high-speed spray image processing — the pipeline automates their core billable task
Enablers (Picks & Shovels)
- PyTorch / torchvision — Faster R-CNN and transformer components used directly
- High-speed camera vendors such as Photron and Phantom — hardware generating the shadowgraphy data
- Synthetic data generation pipelines built on domain-specific image augmentation tools
- Feature Pyramid Networks (FPN) — the multi-scale detection architecture enabling small-droplet detection
Private Watchlist
- Spraying Systems Co. (private) — industrial spray nozzle manufacturer that could use automated breakup analysis
- Sentient Science (private) — physics-informed simulation and diagnostics for industrial components
Resources
The Paper
The disintegration of liquid sheets into ligaments and droplets involves highly transient, multi-scale dynamics that are difficult to quantify from high-speed shadowgraphy images. Identifying droplets, ligaments, and blobs formed during breakup, along with tracking across frames, is essential for spray analysis. However, conventional multi-object tracking frameworks impose strict one-to-one temporal associations and cannot represent one-to-many fragmentation events. In this study, we present a two-stage deep learning framework for object detection and temporal relationship modeling across frames. The framework captures ligament deformation, fragmentation, and parent-child lineage during liquid sheet disintegration. In the first stage, a Faster R-CNN with a ResNet-50 backbone and Feature Pyramid Network detects and classifies ligaments and droplets in high-speed shadowgraphy recordings of an impinging Carbopol gel jet. A morphology-preserving synthetic data generation strategy augments the training set without introducing physically implausible configurations, achieving a held-out F1 score of up to 0.872 across fourteen original-to-synthetic configurations. In the second stage, a Transformer-augmented multilayer perceptron classifies inter-frame associations into continuation, fragmentation (one-to-many), and non-association using physics-informed geometric features. Despite severe class imbalance, the model achieves 86.1% accuracy, 93.2% precision, and perfect recall (1.00) for fragmentation events. Together, the framework enables automated reconstruction of fragmentation trees, preservation of parent-child lineage, and extraction of breakup statistics such as fragment multiplicity and droplet size distributions. By explicitly identifying children droplets formed from ligament fragmentation, the framework provides automated analysis of the primary atomization mode.