Short student presentations

Student presentations

Dibyajyoti Chakraborty

College of Information Sciences and Technology

Application of machine learning in non-Newtonian flows

With the abundance of available data and advancements in machine learning, there is a growing need to shift toward data-driven and optimization-based approaches for solving problems that were traditionally solved using CFD techniques. Neural networks, especially, offer great potential in capturing the chaotic nature of dynamic phenomena. On the other hand, established physical laws serve as benchmarks of knowledge, validating the results obtained through computational and data-driven methods. In this work, we have implemented the latest Physics Informed Neural Networks (PINN) to solve non-Newtonian flows and compared their accuracies with theoretical/numerical solutions and available experimental data. Additionally, we have used numerical simulations to generate data for non-Newtonian flow over spheres under a range of operational parameters. We have shown how data from numerical experiments can be used to predict properties, like drag, which would otherwise have much higher computational costs. Finally, we have demonstrated that machine learning could be used to optimize the shape of objects in non-Newtonian flows without the requirement for performing numerical simulations over the entire domain of different shape parameters.

Ashbell Abraham

Department of Physics

Reversibility, path-dependence, and memory of a creeping triple-phase contact line

The contact line around a water drop on a horizontal surface has an irregular shape that does not relax to equilibrium, revealing the disorder of the solid substrate beneath. We show that the contact line has a detailed memory of its history of motion on the surface. To form a memory, we start with an initial volume of water and “train” the contact line with slow imbibition and drainage cycles at constant volume amplitude until its motion becomes periodic. Reducing the volume amplitude drastically changes the shape of the contact line when it returns to its starting volume, but when driven again with only one cycle at the training amplitude, it transitions back to its steady state. Driving above the training amplitude erases the memory making the steady state inaccessible. This behavior is reminiscent of return-point memory, a phenomenon best known in ferromagnets. Return-point memory, and the evolution to steady state, can give insight on how contact line hysteresis and the memory of its motion arise, offer a framework to study its reversible-irreversible transitions, and provide a comparison of this system to others that are far from equilibrium.

Shyam Nair

Department of Mechanical Engineering

The hydrodynamic properties of unconventional surface roughness

Surface roughness is characterized by its statistics such as the average height, k_a, the root-mean-square, k_rms, the skewness, S_k, etc. A survey of the roughness in the existing literature shows limited work on roughness with low S_k and high k_rms/k_a, which is the focus of this talk. By controlling the spacing, height, and arrangement of rectangular protrusions and pits, such rough walls are constructed. Direct numerical simulations (DNSs) are conducted to study their hydrodynamic properties. The data suggests that Sk has a minor effect on the equivalent sandgrain roughness height (k_s) as pits, that contribute to negative S_k, do not significantly contribute to drag. Furthermore, the roughness arrangement, which is not captured by single-point roughness statistics, has a significant effect on k_s. Attempts in developing correlation-type rough-wall models show that the predictive power of a rough-wall model depends on how well the input space distinguishes the rough walls in the calibration dataset. Moreover, the importance of roughness statistics depends on the rough walls under consideration. These findings suggest that a universal rough-wall model might not be possible, and models for specific types of roughness might be more practical.

Ke Zhou

Department of Mechanical Engineering

Reconstructing complex flows from inertial Lagrangian particle tracks

Particle tracking velocimetry (PTV) is widely used to reconstruct 4D flow states from Lagrangian particle trajectories, a.k.a. “tracks”, assuming that particles faithfully follow the flow. However, particles can lag the flow or travel ballistically due to rapid acceleration, large temperature gradients, strong body forces, etc., complicating the interpretation of PTV data. We report a novel method to simultaneously reconstruct unsteady flow states and individual particle properties (e.g., size, density, effective response time, ...) from inertial tracks. To do this, we use a neural-implicit particle transport model to predict PTV tracks as a function of estimated flow states and particle properties. The flow states and tracks are parameterized by physics-informed neural networks (or similar). Optimizing an objective loss comprising a PTV data loss, Navier–Stokes residuals, and particle transport residuals yields tracks that match the data, physically-plausible flow states, and estimates of the unknown particle properties. We demonstrate this approach using synthetic tracks of inertial particles carried by laminar and turbulent flows. To the best of our knowledge, we report the first unsteady flow reconstructions from inertial tracks as well as implicit PTV-based particle sizing.