Session L28: Low-Order Modeling: Applications

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.   

Nonlinear parametric models of viscoelastic fluid flows

Reduced-order models have been widely adopted in fluid mechanics, particularly in the context of Newtonian fluid flows. These models offer the ability to predict complex dynamics, such as instabilities and oscillations, at a considerably reduced computational cost. 

 In contrast, the reduced-order modeling of non-Newtonian viscoelastic fluid flows is relatively unstudied.  

 This work explores the use of the sparse identification of nonlinear dynamics (SINDy) algorithm to develop interpretable reduced-order models for a broad class of  viscoelastic flows. 

 In particular, we explore a benchmark oscillatory viscoelastic flow on the four-roll mill geometry using the classical Oldroyd-B fluid. 

 This flow exemplifies many canonical challenges associated with non-Newtonian flows, including transitions, asymmetries, instabilities, and bifurcations arising from the interplay of viscous and elastic forces, all of which require expensive computations to resolve fast timescales and long transients.  

 First, we demonstrate the effectiveness of our data-driven surrogate model in predicting the transient evolution on a simplified representation of the dynamical system. We then describe the ability of the reduced-order model to accurately reconstruct spatial flow field in a basis obtained via proper orthogonal decomposition. 

 Finally, we develop a fully parametric, nonlinear model that captures the dominant variations of the dynamics with the relevant nondimensional Weissenberg number. 

 This work illustrates the potential to reduce computational costs and improve design, optimization, and control of a large class of non-Newtonian fluid flows with modern machine learning and reduced-order modeling techniques.   

Using Artificial Intelligence for Transient Heat Transfer

During hypersonic re-entry, heat transfer throughout a vehicle is modeled as a transient time-dependent problem due to the constant deformation of the vehicle from aero-heating effects. Traditional numerical methods, including the finite difference method, have already been widely successful in modeling these transient heat transfer problems. Machine learning frameworks have also been recently proposed to solve problems in dynamic environments, and machine learning algorithms have been applied to stress simulations as quicker alternatives that produce comparable accuracy. To improve the simulation wall-time, this study examined the possible use of machine learning to emulate the finite difference method solver on the 2D heat equation. To generate test data the finite difference method was used to solve the 2D heat equation and generate test data usable by a neural network. Multiple machine learning models were then trained using this test data, and the results of each method were compared amongst themselves, and the data generated by the finite difference method. The feasibility of applying machine learning models to certain problems and whether machine learning models can serve as an alternative or improve current methods was also assessed.   

Estimating thermofluid system parameters using a Markov chain Monte Carlo method, with an example of oscillating heat pipes

When simulating complex thermal and fluid systems with coupled physical models, there are often some model parameters that are not known a priori but substantially influence the observable results. An oscillating heat pipe, or OHP, exemplifies this challenge. This thermal management device consists of a large, dynamic number of liquid slugs interspersed with vapor bubbles inside of a serpentine tube embedded into a metallic plate. In an OHP the characteristics of nucleate boiling and the liquid film on the tube wall adjacent to the bubbles each affect the temperature measured at fixed locations. Data assimilation, and particularly, Bayesian inference, can be used to estimate the unknown parameters and their uncertainties from sufficiently detailed spatio-temporal experimental observations. In this study, we present a data-assimilation framework for an oscillating heat pipe using the Markov chain Monte Carlo (MCMC) method. Critical parameters for nucleate boiling and liquid film dynamics are estimated from temporal experimental temperature observations at fixed locations, rationalized with a relatively low-order physics-based predictive OHP model. Based on the results of several OHPs with different boundary conditions and two different working fluids, we report progress in building adaptive prediction models for thermofluid system performance with limited initial training.   

Data-driven dimensional analysis and modelling of two-phase heat-transfer in small-to-micro tubes

The thermal design of electronic devices requires dissipation of large heat fluxes from small surface areas. This task can be accomplished by using two-phase flow boiling in small-to-micro diameter channels. The design of these compact heat exchangers requires accurate prediction of the two-phase heat transfer coefficient (HTC) as function of the operating conditions. Approaches to predict the HTC consist of empirical and semi-empirical correlations, and machine learning models, all of which involve performing complex nonlinear regressions on experimental data. However, it is challenging to identify the optimal dimensionless groups to be used as input for these regressions due to the complex interactions of key variables in phase-changing flows. To address this problem, we exploit data-driven dimensional analysis. First, starting from the dimensional input parameters, we estimate the active subspace of dimensionless groups, i.e., the dimensionless groups that have the greatest impact on the HTC. To do so, we use a Gaussian process regression (GPR) to model the function describing the HTC and its Jacobian. The GPR is trained with experimental data from the Brunel Two-Phase Flow database. Then, we express these optimal nondimensional groups in terms of products of powers of the Bond number, the Reynolds and Prandtl numbers of the liquid phase, and the reduced pressure and degree of subcooling at the entry to the channels. Our results show that working in nondimensional space is an effective way to avoid data overfitting and to obtain a realistic estimate of the data uncertainty.   

Data-Driven Modeling for Optical Wave Reconstruction in Supersonic and Hypersonic Flows

We extend the methodology of the intensity/slope network (ISNet) (Dubose, et. al. 2020) to allow for flight condition aware reconstruction and correction of distorted optical wavefronts generated by Shack-Hartmann wavefront sensors (SHWFS) on high-speed flight vehicles. The current focus includes optical distortions caused by turbulent fluctuations in super and hypersonic flows over a cavity. Both experimentally and numerically generated distorted wavefronts are used to train the network. The flight conditions are included as an input to allow the network to adjust the reconstruction/correction to match the varying characteristics of the associated turbulence. We also attempt to extend this body of work for application with plenoptic wavefront sensors (PWFS).   

Reduced Order Atmospheric Pollution Modelling using Machine Learning with Proper Orthogonal Decomposition

Atmospheric pollution modelling plays a crucial role, not only in understanding the effects of air pollution on human health and the environment but also in the day-to-day monitoring of air quality. However, the computational cost of high-resolution simulations poses challenges for operational pollution forecasting. To address this problem, we propose an approach using Machine Learning (ML) supervised by numerical simulation results obtained from the WRF-CHEM model with an 800m resolution over a 102*102 spatial domain. The latter includes the city of Marseille and the basins of the port of Marseille-Fos. This enables us to capture the pollution generated by the significant maritime traffic in Marseille. The main pollutants are modeled, including PM10, PM2.5, NO2, SO2, and O3. The high degrees of freedom in the input data, resulting from fine spatial resolution, raise learning difficulties for traditional ML models. To overcome this, we apply Proper Orthogonal Decomposition (POD) to reduce data dimensions while preserving essential information. By capturing dominant variability modes, POD facilitates efficient ML training on the reduced-order data. This hybrid ML-POD model demonstrates promising results, achieving accurate pollution predictions while significantly reducing computational costs.   

Deep Reinforcement Learning for Autonomous Navigation in Complex Flows

In turbulent environments, navigating accurately and efficiently poses significant challenges.

This is especially the case if the only information that can be accessed are the local velocity

gradients. This is the sort of challenging task that a plankton is faced with while drifting in the

oceans, or that of a GPS-denied autonomous underwater vehicle (AUV).

Reinforcement learning has become a popular tool to solve navigation problems

with local flow information. So far, these studies are mostly proofs-of-concept. They

demonstrate that better-than-naive strategies can be learnt purely from experience.

But the next challenge for reinforcement learning is to discover truly smart strategies

that exploit the complex features of the underlying turbulent flow. This requires the

use of state-of-the-art reinforcement learning methods.

Here we explore this path to address the problem of directional navigation, we demonstrate

that DRL consistently matches or beat the performance of approximately optimal strategies

derived analytically. We also discuss the importance of using deep neural networks as opposed

to tabular policies. Overall, this study sheds light on the promising applications of reinforcement

learning in tackling directional navigation problems in turbulent environments.   

Deep Learning for flow field and drag force predictions in dispersed particle flows

Accurate modeling of the flow field and drag forces in fluid-particle systems in Euler-Lagrange and Euler-Euler methods is essential for lab-scale and industry-scale simulations. In this work, we utilize deep learning to predict pressure and velocity fields in multiphase flows with dispersed prolate ellipsoidal particles of aspect ratio 2.5 generated by Particle Resolved Simulations at different Reynolds numbers and solid fractions. A 3D U-Net based model is used, which given the spatial information of an ellipsoid’s neighboring region in the form of a distance function, is trained to generate the pressure and velocity fields. We additionally introduce a novel method to perform super-resolution in the predicted domain near the surface of the particle. The trained model is also tested on its ability to generalize on unseen datasets, and the flow field predictions are then evaluated on their ability to predict particle drag forces. We show that this method of predicting the pressure and velocity field first before performing the downstream task of drag force prediction is more accurate than training the model to predict drag forces directly using the distance function input. We also show the predicted drag forces perform better than correlations being used by CFD-DEM.   

Enhancing Computational Fluid Dynamics Research and Education through AI: The Role of ChatGPT

The advent of artificial intelligence (AI) has brought about transformative changes across various fields, and Computational Fluid Dynamics (CFD) is no exception. This presentation explores the potential of OpenAI's language model, ChatGPT, as a tool for enhancing both research and education in CFD. ChatGPT, trained on a diverse range of internet text, can serve as an interactive tool for explaining complex CFD concepts, answering queries, and providing valuable resources for further learning. When proper prompts are utilized, it can be shown that ChatGPT can help to simplify intricate topics such as turbulence modeling, finite volume method, and boundary conditions, making them more accessible to students. We also demonstrate that ChatGPT can assist in problem-solving, from mathematical issues like the discretization of differential equations to practical challenges in debugging CFD simulation code. Additionally, ChatGPT can be a valuable assistant, offering summaries of research papers, suggesting relevant literature, aiding in data analysis, and even assisting in drafting research papers and reports. We use a newly created plugin called AMReX-Doc to demonstrate how ChatGPT can help users quickly search through document content.