A PINN Methodology for Temperature Field Reconstruction in the PIV Measurement Plane: Case of Rayleigh-Bénard Convection

Marie-Christine Volk, Anne Sergent, Didier Lucor, Michael Mommert, Christian Bauer, Claus Wagner · 2025

Read A PINN Methodology for Temperature Field Reconstruction in the PIV Measurement Plane: Case of Rayleigh-Bénard Convection by Marie-Christine Volk, Anne Sergent, Didier Lucor, Michael Mommert, Christian Bauer, Claus Wagner (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-23801.txt.

Introduction

Marie-Christine Volk and five co-authors rebuild temperature maps of a swirling heated fluid from motion data alone. Their 2025 paper uses physics-informed neural networks, which learn under the rules of the flow equations. The test case is Rayleigh-Bénard convection, where fluid heated from below churns in a cube.

Sensors supplied motion in one vertical plane only, as in real lab setups. The network filled the thin slab around that plane, with sample points packed where data existed and where heat gradients run steep. The rebuilt temperatures matched the true fields from the source simulation. The team then stressed the method with shorter records, noise, and missing edge data. They see the work as a step toward using such networks on real experiments.

Catalog record

CREATORMarie-Christine Volk, Anne Sergent, Didier Lucor, Michael Mommert, Christian Bauer, Claus Wagner
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-23801

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A PINN Methodology for Temperature Field Reconstruction in the PIV Measurement Plane: Case of Rayleigh-Bénard Convection Marie-Christine Volk, Anne Sergent, Didier Lucor, Michael Mommert, Christian Bauer, Claus Wagner Abstract We present a method to infer temperature fields from stereo particle-image velocimetry (PIV) data in turbulent Rayleigh-Bénard convection (RBC) using Physics-informed neural networks (PINNs). The physical setup is a cubic RBC cell with Rayleigh number Ra = 10 7 \mathrm{Ra}=10^{7} and Prandtl number Pr = 0.7 \mathrm{Pr}=$0.7$ . With data only available in a vertical plane A : x = x 0 A:x=x_{0} , the residuals of the governing partial differential equations are minimised in an enclosing 3D domain around A A with thickness δ x \delta_{x} . Dynamic collocation point sampling strategies are used to overcome the lack of 3D labelled information and to optimize the overall convergence of the PINN. In particular, in the out-of-plane direction x x , the collocation points are distributed according to a normal distribution, in order to emphasize the region where data is provided. Along the vertical direction, we leverage meshing information and sample points from a distribution designed based on the grid of a direct numerical simulation (DNS). This approach points greater attention to critical regions, particularly the areas with high temperature gradients within the thermal boundary layers. Using planar three-component velocity data from a DNS, we successfully validate the reconstruction of the temperature fields in the PIV plane. We evaluate the robustness of our method with respect to characteristics of the labelled data used for training: the data time span, the sampling frequency, some noisy data and boundary data omission, aiming to better accommodate the challenges associated with experimental data. Developing PINNs on controlled simulation data is a crucial step toward their effective deployment on experimental data. The key is to systematically introduce noise, gaps, and uncertainties in simulated data to mimic real-world conditions and ensure robust generalization. \geometry textwidth=0.8textheight=0.9 1 Introduction Inferring the temperature in thermally-driven flows is highly relevant in fundamental studies and in various technical applications that require precise analysis and control of heat transfer, such as the design of interior ventilation systems in aircraft or train cabins, to optimize air circulation and ensure thermal comfort [ 1 ] . Especially when dealing with such complex technical applications, experiments are crucial to complement, validate or extend theory and simulation. The flow dynamics can be captured within temporally and spatially highly-resolved three-dimensional (3D) velocity measurements using tomographic particle image velocimetry (tomo-PIV) or particle tracking velocimetry (PTV) [ 2 , 3 ] . These methods have also been applied to the canonical experimental setup for thermally-driven flows: Rayleigh-Bénard convection (RBC) (i.e. [ 4 ] ). The simultaneous measurement of spatially resolved velocity and also temperature fields is also feasible using a more elaborated set-up with temperature-sensitive tracer particles (Thermochromic Liquid Crystals, TLCs). Experiments in RBC were performed for example by Schmeling et al. [ 5 ] , Schiepel et al. and [ 6 ] or Käufer and Cierpka [ 7 ] . These experimental setups, however, can be challenging, since the TLCs usually have a small color-range, are sensitive towards background light and the colour-to-temperature calibration is not trivial [ 8 ] . An alternative method for obtaining temperature trajectories involves the use of neutrally buoyant, wireless temperature sensors, as demonstrated by Gasteuil et al. [ 9 ] . However, the focus of this approach is on obtaining single temperature traces rather than fully resolved spatial temperature fields. Therefore, the establishment of new methods to reconstruct temperature fields is of great interest. A significant amount of work has been done in the past on temperature reconstruction from turbulent flows. One approach for data assimilation or reconstruction in turbulent flows uses methods from Direct Numerical Simulations (DNS) and is called nudging. It was first systematically applied in Clark Di Leoni et al. [ 10 ] to assimilate full-resolved velocity fields out of sparse velocity data. This technique is also promising for multi-field physical problems, as Bauer et al. [ 11 ] demonstrated that applying it to sparse velocity data from turbulent tomo-PIV measurements allows for temperature field reconstruction. One challenge in this approach is that errors in the measured velocity fields can significantly impact the evolution and outcome of the temperature field. Also based on the Navier-Stokes equations, Weiss et al. [ 12 ] suggested recently a method to solve a Poisson equation for the temperature, using velocity fields of turbulent RBC. While this technique is computationally efficient, the challenge arises when applying it to experimental data, as it requires spatially and temporally fully resolved fields. Another method employs machine learning (ML) through the integration of data and partial differential equations (PDEs) within so-called physics-informed neural networks (PINNs). PINNs are first introduced by Raissi et al. [ 13 ] and offer a powerful framework for either forward or inverse problems or assimilation and reconstruction tasks. An example for forward problems is solving the compressible Euler equations ( Wassing et al. [ 14 ] ). PINNs are meshfree, and therefore, compared to classic numerical solvers, highly flexible in applications with no access to fully resolved 3D fields. Several approaches assimilate velocity fields from PTV (e.g. Cai et al. [ 15 ] , Wang et al. [ 16 ] and Steinfurth et al. [ 17 ] ) or from different frameworks (i.e. Eulerian or Lagrangian) for efficient pressure recovery in complex moving systems ( Sundar et al. [ 18 ] ). Recent studies revealed the capabilities of physics-informed neural networks to predict instantaneous temperature fields in turbulent convective flows using 3D velocity fields. Toscano et al. [ 19 ] reconstruct temperature fields from PTV data using a physics-informed Kolmogorov-Arnold architecture. Mommert et al. [ 20 ] explore the influence of activation functions when reconstructing temperature fields out of fully resolved DNS velocity fields and vice versa and Lucor et al. [ 21 ] evaluate the robustness with respect to low resolution in the observation data consisting of sparse temperature data in the bulk and velocity data at the boundaries of the spatial training domain. Although tomo-PIV and PTV are well-established and extensively validated techniques, they remain demanding in terms of equipment. Therefore, in this work, we investigate whether it is feasible to rely only on more standard two-dimensional (2D) velocity measurements for temperature reconstruction. We challenge the PINN to reconstruct temperature fields - in particular with high accuracy in the PIV plane - from stereoscopic velocity data in turbulent RBC by using spatially fully resolved velocity snapshots from DNS (cf. numerical methods in Section 2 ). The use of synthetic data ensures a reliable ground truth, which allows accurate validation of the results. Our PINN approach employs a padding technique, motivated in [ 21 ] , by calculating the PDEs in an enclosing 3D domain around the PIV plane and by employing collocation point sampling strategies to further overcome the lack of 3D information ( Section 3 ). In order to mimic PIV data and to quantify the challenges of not having access to spatially and temporally fully resolved velocity fields, we test the robustness of the temperature reconstruction with respect to the time span of the data and the time increment between snapshots

Read the complete text