CNN-PD: Convolutional Neural Networks for Particle Diffusometry Measurements

Measuring fluid properties (such as viscosity & temperature) on a microscale plays a vital role in areas ranging from fuel cells to pharmaceuticals. These properties can be obtained in real-time by seeding an otherwise featureless fluid with particles & monitoring the particles’ diffusion. However, existing methods such as single-particle tracking & correlation-based measurement are significantly degraded when the fluid is flowing as well as in the presence of diffusion coefficient gradients. This work provides a convolutional neural network (CNN) based alternative for diffusion coefficient measurement that averages PIV-style interrogation regions from successive frames. The averaged images used for training the CNNs can have an arbitrary number of frames (n > 1). The results show that with only two-frame averages, R² values (about 0.99) when there was no flow or a known uniform flow. Moreover, the CNNs maintained high R² (about 0.96) without retraining, even when the underlying frames had a diffusion coefficient gradient. Additionally, similar R² values were obtained for arbitrary flows with four-frame averaged images. It shows that the CNNs learn useful representations of diffusion coefficient from temporally averaged particle images & can generalize it to unseen cases.