Application of Machine Learning for the Identification of 2D Colloidal Assemblies: A Case Study on Particles of Distinct Shapes
This work addresses the problem of identifying colloidal monolayer assemblies using particles of various shapes (two-dimensional coatings): spheres, ellipsoids, cuboids, and rods. The following classification of assemblies is considered: isolated particles, dimers, chains, clusters, and loops. The YOLO model was chosen as the identification method. Synthetic datasets were prepared for each of the four particle shapes to train the models. The paper discusses the application of models trained on synthetic data to experimental images. An analysis was carried out on the feasibility of using such models for recognizing configurations in real images. While recognition on artificial images is nearly perfect, tests on experimental images showed a significant deviation. The average error across all particle types was 43.1%, but a considerable spread in values is observed: from 20% for spheres to 58.5% for cuboids, indicating the algorithm's selective sensitivity to object geometry. The created datasets and trained models are freely available for use. The corresponding modules have been integrated into the previously developed information system (https://isanm.space/). To further improve prediction results, it is necessary to prepare datasets based on experimental images.