Classification of time-variant transcritical droplets with attention-enhanced deep spatiotemporal learning

The transcritical mixing of liquid fuel sprays is a process characterised by a fuel droplet's transition from a classical evaporation state to a state of diffusive mixing. Although we previously proposed a phenomenological model identifying distinct mixing regimes (classical evaporation, transitional mixing, and diffusive mixing), analysing such large image datasets still requires significant human intervention. This is primarily due to these mixing regimes being identified through temporal criteria (i.e. evolution of droplet shape in time), which is particularly challenging to automate using traditional image processing algorithms.

To address this issue, we designed a deep spatiotemporal learning algorithm trained on human-annotated frames of synthetic transcritical droplets, inspired by high-speed long-distance microscopy videos. In this paper we present our two-stage detection and classification pipeline, where a multiple-object tracking (MOT) algorithm based on YOLOv11 and an integrated BoT-SORT tracking layer initially detect moving droplets and isolate them at the object level. Then, a novel residual convolutional neural network and bidirectional long short-term memory network with a temporal attention module (CNN-BiLSTM-TAM) is proposed to classify the mixing regimes using the extracted object-level droplet images. Our algorithm is designed to learn both the rich visual characteristics of the droplets and their time-based evolution, using spatial and temporal attention to capture the most informative frames in the droplet image sequences.

We provide robust empirical validation of our work through attention map visualisation and performance comparison with two state-of-the-art image classifiers, showcasing improvements in precision, recall, and F1 score of +50%, +31%, +40% and +81%, +56%, +72%.