Authors:

Shayak Chatterjee

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Particle diffusometry (PD) has emerged as a useful method of finding the diffusion coefficient of a particle in a microfluidic setup with utilities in pathogen detection, pharmaceuticals and fuel cells. Conventional methods used to identify diffusion coefficient are single particle tracing and correlation based methods. Studies have been conducted to investigate application of neural networks for PD and results show that neural networks perform better than the aforementioned conventional methods. However, the current training approach relies on simulated data which utilizes a random walk model without considering essential physics, such as particle- particle and particle-fluid interactions. Our objective is to enhance the training dataset by incorporating improved physics. We have used computer simulations with particle fluid interactions to generate the physics- based dataset. Initial benchmarking involves comparing the physics-based dataset with the original random walk dataset (non-physics). Evaluation metrics include the mean absolute displacement of particles between timesteps and their positions at t = 0s. Furthermore, we will assess the performance (metrics such as mean absolute distance covered by particles and R^2 value) of the neural network trained on the physics-based dataset compared to the non-physics-trained network and conventional techniques mentioned before. This study aims to improve the physics fidelity of the training dataset and enhance the accuracy of the neural network in determining diffusion coefficients.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Shayak Chatterjee