Neural Networks for Particle Diffusometry Measurements in the Presence of Flow and with Defocused Particles

Diffusion is an inherent phenomenon in fluid systems, and it can be optically measured by introducing tracer particles into an otherwise featureless medium and tracking their motion under a microscope. Traditional particle-based algorithms for estimating diffusion coefficients, however, exhibit several limitations—especially when dealing with fluid flow or defocused particles. This study proposes the use of Convolutional Neural Networks (CNNs) as a robust alternative for predicting diffusion coefficients from PIV-style image crops, even under challenging, real-world conditions. The networks were trained, validated, and tested on datasets containing Gaussian-shaped or defocused particles subjected to arbitrary flow conditions. The results show that the CNNs have a low Mean Absolute Error (MAE) of 0.09 um²/s and 0.07 um²/s between the true and predicted diffusion coefficient values for the dataset with Gaussian-shaped and defocused particles, respectively. The performance of the CNNs was also benchmarked against four conventional algorithms on the simulated datasets. The results show that the CNNs outperform conventional methods when the particles are defocused. Finally, the outputs of CNNs were compared against those of conventional algorithms on experimental datasets, resulting in uncertainty ranging from 0.19 μm²/s to 0.47 μm²/s. Hence, the study utilized neural networks to reliably predict diffusion coefficients from complex particle datasets where the conventional algorithms fail.