Session: 11-03-01: Artificial Intelligence, Machine Learning and Data Science for Thermal Processes, Heat Transfer and Energy Systems
Paper Number: 141449
141449 - Reconstructing Invisible Fluid Fields of a Transient Natural Convection System Using a Physics-Informed Neural Network
The ability to reconstruct a fluid field given sparse, often incomplete data, is a major area of interest and a challenging inverse problem found in various science and engineering applications. Complete knowledge is usually required to employ traditional methods of solving flow fields, such as computational fluid dynamics (CFD). With unknown or uncertain boundary conditions, the governing equations are likely to be ill-posed as an underdetermined system. The ability to reconstruct fields in the entire domain, given only sparse data points and limited knowledge of the domain, will allow for complete understanding of fields. For example, many thermofluidic industrial systems only allow for point measurements of temperature and velocity at limited locations. Thus, to analyze the whole system using CFD requires data assimilation and interpolation. Advances in machine learning have explored alternative methods of reconstructing fields using sparse data. For instance, studies using physics-informed neural networks (PINN) to solve thermofluidic problems have gained traction in recent years and work has been done to use PINNs to reconstruct flow fields. A typical PINN follows the idea of minimizing losses coming from a fully connected network, in which the loss functions impose constraints such as initial and boundary conditions. The key interest in such a network is that it is also able to constrain the network to satisfy given physical conditions, often governed by partial differential equations (PDEs). Because of the ability of a neural network to perform backpropagation, the PDEs can be solved exactly without complicated discretization schemes such as the Galerkin method, which is commonly used in CFDs.
This work aims to reconstruct the transient natural convection within a 2D square domain. Sparse temperature fields will be used to infer pressure and velocity fields. Synthetic temperature data is generated with Ansys Fluent and fields are exported for ground truth reference at 0.1 second intervals. Sparse temperature data and wall boundary conditions are the only constraints. We aim to explore the effects of transience in PINN accuracy, in which degrees of freedom (DoF) and entropy in fields are examined and correlated to the performance of the network.
Different DoFs and entropy of unsteady behavior may exist under various timescales, which may not be represented uniformly by a single PINN. For example, the incipient phase will include large time derivatives (equivalently large DoF) while the quasi-steady phase will essentially have zero time derivatives (equivalently small DoF). These phases may require individual PINNs that operate with different levels of constraints. Furthermore, this work uses water with a larger Prandtl number (Pr) of 4.8 as compared to air with a Pr of 0.7, which adds an extra scale into the governing equations and numerical complexity.
A method of quantifying transience was proposed. Calculation of the absolute difference (AD) of fields between timestep gives insight into the gradients in the temporal direction, in which a larger gradient indicates faster flow development. It was observed that the difference between starting frame and ending frame errors correlated with the magnitude of the AD. Taking inspiration from information theory, another metric was proposed to quantify field complexity. Here, the entropy of each field is calculated for each timestep, showing that low entropy regions correlated with low reconstruction accuracy.
*NSF Grant No. 2053413
Presenting Author: Nagahiro Ohashi Arizona State University
Presenting Author Biography: Nagahiro is a PhD student in mechanical and aerospace engineering at the Arizona State University. He received his BS in 2021 and MS in 2023 in mechanical engineering from the University of Hawaii at Manoa. His research interest focuses on using deep learning techniques to derive temperature distributions via a network trained with experimental data.
Authors:
Nagahiro Ohashi Arizona State UniversityNam Nguyen Arizona State University
Leslie Hwang Arizona State University
Beomjin Kwon Arizona State University
Reconstructing Invisible Fluid Fields of a Transient Natural Convection System Using a Physics-Informed Neural Network
Paper Type
Technical Presentation