Atmospheric droplet detection and classification with self-mixing interferometry and neural networks

Self-mixing interferometry presents obvious advantages (simplicity, energy consumption, and robustness) for airborne operation. We design an experimental system able to model several atmospheric conditions in terms of water droplet sizes and use a self-mixing interferometer to probe these droplets. We design a small (embeddable) neural network and train it to classify the different types of self-mixing signals in the presence of different types of droplets in a laboratory context. Our results open the way for AI-enabled self-mixing sensors for atmospheric condition monitoring and early warnings of icing conditions. We obtain more than 93% discrimination accuracy in this laboratory setting, but further validation of the approach for airborne operation requires testing in high-speed conditions.