YOLOv8 trained on synthetic images identifies 2D colloidal assemblies at 97–99% on synthetic data but shows a 43.1% average error on real micrographs, from 20% for spheres to 58.5% for cuboids

The classification of colloidal assemblies is extended from spherical particles to four shapes (ellipsoids, rods, cuboids, spheres) and treats loops as a separate class, giving five categories: isolated particles, dimers, chains, loops, and clusters. Earlier work with this classification considered only spherical particles and did not identify loops as a separate group, even though closed structures contribute to gel elasticity; this work applies the scheme uniformly across non-spherical particles and provides a framework for automated recognition. Each category is defined explicitly through neighbor relations (two particles are neighbors if their surfaces touch) and illustrated; four datasets were created, each containing 135–155 images, augmented by rotation and added noise, and split into training, validation, and test sets.

Replacing bounding boxes with polygonal instance segmentation correctly separates closely located particles and avoids incorrectly merging independent particles into a single region. The authors' previous study found bounding boxes insufficiently accurate for complex-shaped assemblies, especially intertwined chains, clusters, and loops; here YOLOv8-seg with polygonal annotation is used to delineate each object's boundary precisely. The method section states that polygons handle partial overlap and complex morphology; dataset records store a class number followed by normalized polygon vertex coordinates, with classes 0 chain, 1 cluster, 2 dimer, 3 isolated particle, 4 loop.

On synthetic data the four models reach about 97–99% recognition accuracy, but performance drops substantially on experimental images, with an average relative error of 43.1% that grows with particle geometric complexity. This quantifies the transferability boundary of training purely on synthetic data and provides a shape-ordered error spectrum: 20% for spheres, 44.7% for ellipsoids, 49.2% for rods, and 58.5% for cuboids. On synthetic test sets, sphere Precision and Recall are both 1, while ellipsoids, cuboids, and rods remain in the 0.94–1 range; on experimental data, spheres give Precisionc = 0.73 and Recallc = 0.88, rods drop to Precisionc = 0.31, and cuboids give Recallc = 0.38; error is computed as δ = |Xe − Xs|/Xs × 100% and averaged over four metrics.

The error structure is imbalanced: precision degrades more than recall, indicating that the main problem is false positives from background noise, debris, and labels rather than missed detections. On this basis the authors argue a single metric is insufficient, switch to soft precision and soft recall where probability overlap occurs, and extend evaluation from whole assemblies to individual particles in the images. Average relative error of precision is 52% for configurations and 47% for particles, versus 36.85% and 36.35% for recall; for rods, configuration precision drops by 67% while recall drops by only 35%; the soft metrics follow Fränti and Mariescu-Istodor.

The framework targets automated identification of two-dimensional monolayer colloidal assemblies under a simplified neighbor definition in which particles are neighbors when their surfaces touch, covering four shapes (spheres, ellipsoids, cuboids, rods) and five assembly classes (isolated particles, dimers, chains, loops, clusters). It serves settings that need high-throughput counting of structural motifs, for example supplying experimental structural data for percolation and master-equation kinetic models of gelation. The datasets and training scripts are released under a gPL-3.0 licence and the corresponding modules are integrated into the isanm.space information system, so engineers and researchers in related fields can reuse them or use them as a baseline. The authors state that improving real-image performance requires building a new training dataset from experimental micrographs, suggest comparing future models with theirs on specific images, and note mixtures of particles of different shapes as an interesting next case.

The experimental evaluation rests on only 40 experimental images (ten per shape), so the per-shape error values should be read as observations on this sample rather than universal constants. The cuboid dataset was the only one of the four showing pronounced overfitting, which led the authors to limit training to 35 epochs and to judge that optimal model quality occurs around epoch 30–50; how that choice affects the final metrics remains worth watching. Particle truncation (loss of particle fragments) is explicitly described as not completely understood, and a "500 nm" scale label was misrecognized as a chain, showing that non-particle elements such as scale bars and labels can disturb counting. Real particle shapes often deviate from the ideal geometric models, particles labeled ellipsoidal may in practice resemble rods, rod-like particles are sometimes curved, and a cuboid is a special case of a rod while a sphere is a special case of an ellipse; how these borderline cases affect classification consistency remains an open question. In addition, this is a fast-parse version in which the confusion matrices, loss curves, and identification example figures (Figs. 2–13) are not provided with the text, so the detailed per-class confusion structure and the overfitting curves cannot be checked.