Decrypting the temperature field in flow boiling with latent diffusion models
UngJin Na a , JunYoung Seo b , Taeil Kim a , ByongGuk Jeon c , HangJin Jo a, b, · 2025
Read Decrypting the temperature field in flow boiling with latent diffusion models by UngJin Na a , JunYoung Seo b , Taeil Kim a , ByongGuk Jeon c , HangJin Jo a, b, (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-2501-16510.txt.
Introduction
UngJin Na and co-authors teach an image model to paint heat maps of boiling flow from bubble maps in a 2025 paper. Heat fields in boiling are hard to read and costly to fake with full math. The model learns from a bank of worked-out flow cases in two stages: first a codebook squeezes the maps, then a cleaning net fills in the heat.
The made heat fields match the true ones well at broad and mid scales. Fine scales still drift, the team notes. The shortcut cuts the cost of full runs and sharpens how lab readings are tuned. Next steps aim at finer churn and wider boiling states.
Catalog record
| CREATOR | UngJin Na a , JunYoung Seo b , Taeil Kim a , ByongGuk Jeon c , HangJin Jo a, b, |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-16510
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Decrypting the temperature field in flow boiling with latent diffusion models UngJin Na a , JunYoung Seo b , Taeil Kim a , ByongGuk Jeon c , HangJin Jo a, b, Abstract This paper presents an innovative method using Latent Diffusion Models (LDMs) to generate temperature fields from phase indicator maps. By leveraging the BubbleML dataset from numerical simulations, the LDM translates phase field data into corresponding temperature distributions through a two-stage training process involving a vector-quantized variational autoencoder (VQVAE) and a denoising autoencoder. The resulting model effectively reconstructs complex temperature fields at interfaces. Spectral analysis indicates a high degree of agreement with ground truth data in the low to mid wavenumber ranges, even though some inconsistencies are observed at higher wavenumbers, suggesting areas for further enhancement. This machine learning approach significantly reduces the computational burden of traditional simulations and improves the precision of experimental calibration methods. Future work will focus on refining the model’s ability to represent small-scale turbulence and expanding its applicability to a broader range of boiling conditions. a Department of Mechanical Engineering, Pohang University of Science and Technology (POSTECH), Pohang, Republic of Korea b Division of Advanced Nuclear Engineering, Pohang University of Science and Technology (POSTECH), Pohang, Republic of Korea c Korea Atomic Energy Research Institute, Daejeon, Republic of Korea K eywords Flow Boiling \cdot Latent Diffusion Models 1 Introduction Flow boiling plays an important role in enhancing the performance of thermal management systems, including refrigeration, microelectronics cooling, nuclear power plants, and nuclear fission reactors [ 1 , 2 ] . This phenomenon involves a fluid absorbing heat and undergoing a phase change from liquid to vapor, while supplied with the advection of the bulk flow, significantly boosting the heat transfer efficiency through the utilization of latent heat. The initiation of the phase change is known as the onset of nucleate boiling (ONB) [ 3 ] . However, when the liquid fails to rewet the surface, the surface becomes entirely covered by a vapor layer, leading to a significant reduction in heat transfer efficiency. This phenomenon is known as the departure from nucleate boiling (DNB) [ 4 ] . The heat transfer process between the ONB and the DNB points can be described using the RPI wall boiling model [ 5 ] . This model breaks down the total heat flux from the wall to the working fluid into three components: convective heat flux ( q_{c} ), quenching heat flux ( q_{q} ), and evaporation heat flux ( q_{e} ). The convective heat flux accounts for the heat transfer between the liquid phase and the heated wall surface not covered by bubbles. The quenching heat flux represents the cyclic averaged transient heat transfer as liquid refills the wall vicinity after bubble detachment. The evaporative heat flux is associated with the latent heat carried away by departing bubbles. To estimate the heat transfer, visualization methods such as side-view shadowgraphs, total reflection, and infrared thermography have been utilized as major tools for collecting information on bubble dynamics [ 6 , 7 ] . Additionally, Laser-Induced Fluorescence (LIF) techniques have been adopted to experimentally measure the spatial temperature of the fluids [ 8 ] . However, accurately estimating the actual temperature field remains extremely challenging. Furthermore, detailed computational models for subcooled boiling flow, such as the Eulerian-Eulerian framework, involve establishing mass, momentum, and energy conservation equations for each phase separately, which requires tremendous resources to calculate [ 9 ] . Recent advancements in machine learning, particularly in convolutional neural networks (CNNs), offer hope in addressing these challenges [ 10 ] . CNNs have shown proficiency in capturing visual features, suggesting the potential to simplify complex experimental setups and data extraction processes [ 11 ] . Also, generative models, especially Generative Adversarial Networks (GANs), have shown promise in image processing tasks by learning the style of input data and generating corresponding images [ 12 ] . However, GANs face challenges such as mode collapse and instability during training, limiting their effectiveness in modeling complex distributions [ 13 ] . Addressing these limitations, denoising diffusion probabilistic models (DDPMs) have emerged as a robust alternative for high-resolution image synthesis [ 14 , 15 ] . These models, built from a hierarchy of denoising autoencoders, have achieved impressive results in image synthesis without mode collapse and training instabilities, setting new benchmarks in conditional image generation and super-resolution tasks [ 16 ] . Latent Diffusion Models (LDMs) operate by first training an autoencoder to create a lower-dimensional representation of the data, reducing the dimensionality of the input to address computational inefficiencies, and then performing the training and sampling [ 17 ] . This study presents a pioneering approach that utilizes LDMs for synthesizing temperature data in a flow boiling system. The model is trained to map phase field data to corresponding temperature field data. By integrating advanced conditioning mechanisms, the model learns and generates outputs even without explicitly known transfer functions between the input and output fields. The autoencoder produces understandable latents from ground truth information, which are later used for the denoising process. Our method provides a robust framework for generating high-fidelity thermal information of the field, which is critical for understanding complex thermal-hydraulic interactions during flow boiling processes. This capability makes LDMs a powerful tool for analyzing digitized bubble information and provides access to high-resolution information to a broader range of researchers and applications. 2 Implementation 2.1 Dataset Utilization This study utilizes the numerical simulation results of flow boiling from the BubbleML dataset, acquired by the Flash-X software [ 18 , 19 ] . In particular, the phase field indicators and the temperature field data are utilized. To introduce this dataset, the simulation was conducted by solving the mass, momentum, and energy conservation equations for incompressible flow. A single-fluid approach is adopted, with variable properties based on the location of the interface. The momentum and energy equations are non-dimensionalized with reference quantities from the liquid phase defining the Reynolds number, Prandtl number, and Froude number. Temperature is scaled using (T-T_{\text{wall}})/(T_{\text{wall}}-T_{\text{bulk}}) . The liquid-vapor interface is tracked using a level-set advection equation. The fluid used is FC-72, known for its stability and insulation properties. The simulation has a domain resolution of n_{x}=1344 , n_{y}=160 , for the domain size of l_{x}=29.4\,\text{mm} , l_{y}=3.5\,\text{mm} . The non-dimensional constants for the fluid can be found in [ 18 ] . Simulation outputs are stored in HDF5 files with variables such as velocities, temperature, and signed distance function. 2.2 Implementation of the LDMs In this study, we utilized LDMs for image-to-image translation. The key advantage of LDMs lies in their ability to separate the training process into two distinct phases: perceptual compression and semantic compression. In the first phase, the perceptual autoencoder eliminates high-frequency details while retaining essential perceptual features. In the second phase, the diffusion model learns the semantic composition of the data within this compressed latent space, derived from the autoencoder, significantly reducing computational complexity while maintaining high-fidelity image generation. To elaborate, the operational procedure of LDMs initiates with the encoding of the original image x_{0} into a latent representation z_{0} using the encoder function E , such that z_{0}=E(x_{0}) . Here, we used the vector-quantized variational autoencoder (VQVAE) for the effective compression of