Register now After registration you will be able to apply for this opportunity online.
Physics-informed machine learning in microfluidics
Understanding the distribution and mechanics of velocity and pressure within microaneurysms is crucial for controlling microrobots navigating through them. Traditional methods for velocity and pressure measurement in microchannels, such as particle image velocimetry (PIV) and numerical simulations based on fluidic physics laws, suffer from high computational demands and inability to operate in real-time. Moreover, pure image methods struggle with near-wall regions lacking visible particles. Leveraging recent advancements in machine learning, particularly convolutional neural networks (CNNs), this project proposes a novel approach - a physics-informed CNN integrated with Navier-Stokes equations and optical flow equations. This CNN aims to accurately predict velocity and pressure profiles in microchannel flows in real-time using only flow images and essential physical parameters. The network architecture comprises an encoder-decoder structure with seven convolutional layers, incorporating down-sampling and up-sampling layers. The final output layer produces three channels representing horizontal velocity, vertical velocity, and pressure. Additionally, a physics-informed loss function, incorporating dimensionless Navier-Stokes equation residuals and optical flow equation residuals, enhances the model's performance by integrating knowledge of fluid dynamics and computer vision. This approach represents a promising advancement towards achieving real-time, high-accuracy prediction of velocity and pressure fields in microchannel flows, with potential applications in microrobotics and microfluidics.
This study explores the potential of physics-informed machine learning (ML) techniques to predict complex fluid dynamics, specifically focusing on flow patterns generated in microfluidic environments. Using images captured with a high-speed camera, depicting various flow patterns resulting from disturbed flow profiles induced by microparticle injection, we aim to develop a deep neural network (DNN) capable of predicting both the next frame in the sequence and providing insights into pressure and velocity distributions.
The proposed DNN leverages physics-informed ML methodologies, combining deep learning with operator learning techniques, to model and predict intricate fluid dynamics phenomena. Notably, our approach extends beyond 2D predictions, aiming to represent velocity and pressure fields in three dimensions, thereby offering a comprehensive understanding of the fluid behavior.
By validating our model's performance against experimental data from microfluidic setups, we aspire to transition from the laboratory to real-world applications, particularly in biomedical engineering contexts. The ability to accurately predict fluid dynamics in vivo holds significant promise for enhancing our understanding of biological processes and improving diagnostic and therapeutic techniques.
Moreover, as physics-informed ML methods have demonstrated success in simplified Navier-Stokes equations and turbulence modeling, our study seeks to broaden the application of these techniques. We envision deploying our AI-driven approach to address broader challenges in climate modeling, sustainability initiatives, and other interdisciplinary fields, thereby contributing to advancements in science and engineering with profound societal impacts.
This study explores the potential of physics-informed machine learning (ML) techniques to predict complex fluid dynamics, specifically focusing on flow patterns generated in microfluidic environments. Using images captured with a high-speed camera, depicting various flow patterns resulting from disturbed flow profiles induced by microparticle injection, we aim to develop a deep neural network (DNN) capable of predicting both the next frame in the sequence and providing insights into pressure and velocity distributions.
The proposed DNN leverages physics-informed ML methodologies, combining deep learning with operator learning techniques, to model and predict intricate fluid dynamics phenomena. Notably, our approach extends beyond 2D predictions, aiming to represent velocity and pressure fields in three dimensions, thereby offering a comprehensive understanding of the fluid behavior.
By validating our model's performance against experimental data from microfluidic setups, we aspire to transition from the laboratory to real-world applications, particularly in biomedical engineering contexts. The ability to accurately predict fluid dynamics in vivo holds significant promise for enhancing our understanding of biological processes and improving diagnostic and therapeutic techniques.
Moreover, as physics-informed ML methods have demonstrated success in simplified Navier-Stokes equations and turbulence modeling, our study seeks to broaden the application of these techniques. We envision deploying our AI-driven approach to address broader challenges in climate modeling, sustainability initiatives, and other interdisciplinary fields, thereby contributing to advancements in science and engineering with profound societal impacts.
The following tasks are included in this project:
- Running deep learning code in PyTorch
- Image processing and synthesis
- Computer vision
- Modeling fluid dynamics with machine learning
Experience with coding in python is necessary and knowledge of machine learning and fluid dynamics is desirable.
The following tasks are included in this project:
- Running deep learning code in PyTorch - Image processing and synthesis - Computer vision - Modeling fluid dynamics with machine learning
Experience with coding in python is necessary and knowledge of machine learning and fluid dynamics is desirable.
Please send your CV and transcript of records to Mahmoud Medany: mmedany@ethz.ch This project will be an artificial intelligence collaboration between Prof. Daniel Ahmed and IBM Research, Zurich. Please send emails for the application.
Acoustic Robotics Systems Lab (ARSL). Department of Mechanical and Process Engineering (D-MAVT). RSA G 324, Säumerstrasse 4, 8803 Rüschlikon, Switzerland. Website: https://arsl.ethz.ch/
Please send your CV and transcript of records to Mahmoud Medany: mmedany@ethz.ch This project will be an artificial intelligence collaboration between Prof. Daniel Ahmed and IBM Research, Zurich. Please send emails for the application.
Acoustic Robotics Systems Lab (ARSL). Department of Mechanical and Process Engineering (D-MAVT). RSA G 324, Säumerstrasse 4, 8803 Rüschlikon, Switzerland. Website: https://arsl.ethz.ch/