🔬 applied physics

Bubble2Heat: Optical to Thermal Inference in Pool Boiling Using Physics-encoded Generative AI

This paper presents a physics-encoded conditional generative adversarial network framework that leverages simulation data and a specialized preprocessing pipeline to generate high-resolution thermal fields from standard optical and pointwise temperature measurements in pool boiling experiments, effectively bridging the gap between observable multiphase phenomena and underlying heat transfer dynamics.

Original authors: Qianxi Fu, Youngjoon Suh, Xiaojing Zhang, Sanghyeon Chang, Yoonjin Won

Published 2026-02-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qianxi Fu, Youngjoon Suh, Xiaojing Zhang, Sanghyeon Chang, Yoonjin Won

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Invisible" Heat

Imagine you are watching a pot of water boil. You can clearly see the bubbles forming, rising, and popping. That's the visual part. But what you can't see is the temperature map underneath those bubbles. Where is the water hottest? Where is it cooling down?

In the real world, measuring this temperature field is like trying to take a high-speed photo of a ghost. The bubbles move too fast, the water is chaotic, and putting a thermometer in the way would disturb the very thing you are trying to measure. Scientists usually have to rely on computer simulations to guess what the heat looks like, but those simulations are expensive and hard to make match real life perfectly.

The Solution: A "Magic Translator" AI

The researchers built a special AI called Bubble2Heat. Think of this AI as a creative translator or a weather forecaster that has never seen a real storm but has studied millions of perfect, computer-generated storms.

Here is how it works, step-by-step:

1. The Training Phase (Learning from a Video Game)

Since they couldn't get perfect temperature data from real boiling water, they first taught the AI using a computer simulation (a fancy video game of physics).

The Input: They showed the AI thousands of frames of "bubble shapes" (where the bubbles are) and a few simple temperature numbers (like a thermometer reading).
The Output: The AI learned to draw the full, detailed heat map that matches those bubbles.
The Trick: They didn't just show the AI one picture at a time. They showed it a movie clip (a sequence of bubbles moving). This helped the AI understand that a bubble doesn't just appear; it grows, moves, and changes, which helps it predict the heat trail it leaves behind.

2. The "Physics-Encoded" Secret Sauce

Usually, AI just guesses patterns. If you show it a picture of a cat, it guesses "cat." But here, the researchers wanted the AI to understand physics.

They didn't write complex math equations into the AI's brain. Instead, they fed it so many examples of how bubbles and heat actually behave in the simulation that the AI naturally learned the "rules of the game."
They also used a technique called Data Augmentation. Imagine teaching someone to recognize a car by showing them pictures of cars facing left, right, upside down, and mirrored. This stops the AI from getting confused if the camera angle changes in the real world.

3. The Real-World Test (The "Magic Trick")

Once the AI was trained on the computer simulation, they tried it on real boiling water.

The Input: They took high-speed videos of real bubbles and a few simple temperature readings from a thermometer stuck in the heater.
The Output: The AI generated a full, detailed temperature map of the real water, even though it had never "seen" real water before. It successfully translated the visual bubble shapes into invisible heat data.

How Good Was It?

The paper compares the AI's guesses to the "ground truth" (the perfect simulation data) and found:

It's accurate: The AI could predict the temperature with a very small error margin (less than 6% off).
It follows the rules: Even though it was trained on a computer, the heat maps it made for real water followed the correct physical trends. For example, it correctly guessed that the solid heater is the hottest, the vapor (bubbles) is next, and the liquid water is the coolest.
It handles time: Because the AI watched the "movie" of bubbles, it could predict how the heat changed over time, not just in a single frozen moment.

The Limitations (The "Fine Print")

The authors are honest about where the AI struggles:

The "Bubble Behavior" Gap: The AI was trained on simulations where bubbles behaved in a specific, somewhat simple way. In the real world, bubbles can get chaotic and crowded in ways the AI hasn't seen. When the real bubbles acted very differently from the training bubbles, the AI's guesses got a bit wobbly.
The "Zoom" Issue: If you look at the real bubbles from a different distance (zooming in or out), the AI gets confused because the bubble shapes look different than what it learned. It needs the "view" to be similar to its training data to work best.

The Bottom Line

This paper introduces a new tool that lets scientists see the invisible heat in boiling water just by watching the bubbles. It's like having a thermal camera that doesn't need expensive hardware; it just needs a regular camera and a smart AI that learned the rules of boiling from a computer simulation. While it's not perfect yet, it bridges the gap between what we can see (bubbles) and what we need to know (temperature).

Technical Summary: Bubble2Heat: Optical to Thermal Inference in Pool Boiling Using Physics-encoded Generative AI

Problem Statement
Quantitative characterization of multiphase heat transfer in pool boiling is hindered by the difficulty of measuring temperature fields within chaotic, rapidly evolving flow regimes. While computational fluid dynamics (CFD) can provide high-resolution data, replicating complex experimental conditions remains prohibitively difficult. Conversely, existing experimental diagnostics (e.g., microheater arrays, infrared thermography, PLIF) are often constrained by delicate setups, limited spatial/temporal resolution, or an inability to provide dense, field-level thermal data suitable for generalized modeling. There is a critical need for a method to extract rich thermal information from readily accessible measurements, such as high-speed optical recordings and pointwise thermocouple data, without requiring intrusive or advanced sensing techniques.

Methodology
The authors propose a deep learning framework, termed Bubble2Heat, which utilizes a Physics-encoded Conditional Generative Adversarial Network (PECGAN) to infer 2D temperature fields from segmented high-speed optical images and pointwise temperature references.

Data Generation: The framework relies on a hybrid Lattice Boltzmann and Finite Difference (LB-FD) simulation method to generate the ground truth training data. This dataset consists of paired phase contour maps (solid, liquid, vapor) and normalized temperature fields across 9 distinct simulation sets with varying heater configurations and temperatures.
Preprocessing and Alignment:
- Simulation Data: Phase contours are derived by thresholding density fields. Temperature maps are normalized into a modified Jakob number ( $T^*$ ).
- Experimental Data: High-speed images are processed using a pre-trained Mask R-CNN for instance segmentation to isolate vapor regions. Reference temperature maps are manually constructed from thermocouple readings and normalized to match the simulation scale.
Model Architecture: The PECGAN employs a U-Net-like generator and a feedforward discriminator.
- Inputs: The generator receives a stack of phase contour maps (current and $p$ past frames) and a reference temperature map ( $T_{ref}$ ).
- Training Objective: The generator minimizes a weighted combination of temperature loss ( $L_T$ ), gradient loss ( $L_{\nabla T}$ ), and adversarial loss ( $L_{GAN}$ ). The discriminator is trained to distinguish real from generated image pairs.
- Physics Encoding: Instead of hard-coding governing equations (as in PINNs), the model encodes physics implicitly through the training data and the use of sequential inputs (temporal context) to maintain consistency with bubble dynamics.
Data Augmentation: To compensate for the lack of explicit physical constraints and improve robustness, the authors apply horizontal mirroring (perspective invariance) during training. This prevents the model from relying on fixed viewing orientations.

Key Contributions

Framework Development: A novel deep learning pipeline that bridges the gap between observable multiphase phenomena (optical images) and underlying thermal transport (temperature fields) using simulation-trained generative models.
Physics-Encoded Inference: Demonstration that a Conditional GAN (CGAN) trained solely on simulation data can be directly applied to experimental data to infer temperature fields, provided the input data is preprocessed into a shared format (phase masks + reference temperature).
Temporal Context Strategy: Implementation of a many-to-one image-to-image inference scheme (using a sequence of past phase contours) to address issues where bubble contours may vanish or become unstable due to quantization or segmentation errors, thereby enhancing temporal consistency.
Augmentation for Physical Plausibility: Validation that standard data augmentation strategies (e.g., image flipping) effectively enhance the physical plausibility of the inference and reduce directional bias, even in the absence of strict physical constraints.

Results

Simulation Performance: The model ( $G_{p=3, T_{ref}}^+$ ) achieved an average mean absolute error (MAE) of approximately 0.0395 in normalized temperature units for the fluid region. Increasing the number of past phase contours ( $p$ ) generally reduced inference error, particularly for temporal consistency, though it slightly increased gradient error due to local noise.
Physical Trends: The inferred results successfully reproduced global trends observed in the simulation, such as the increase in excess temperature and surface temperature difference with increasing void fraction. The model also correctly maintained the relative temperature ordering: Solid > Vapor > Liquid.
Experimental Application: When applied to experimental data, the model preserved qualitative physical trends, such as the temperature ordering across phases. However, quantitative accuracy was limited by the significant mismatch in thermodynamic conditions and bubble dynamics (e.g., void fraction ranges) between the training simulation data and the experimental test data.
Limitations: The study acknowledges that inference errors increase when bubble behaviors (e.g., nucleation vs. departure patterns) or thermodynamic regimes in the test data deviate significantly from the training distribution. The model struggles with high heat flux regimes where void fractions exceed the range seen in training.

Significance and Claims
The paper claims that deep generative models offer a powerful approach to augment and interpret experimental measurements in complex two-phase systems where direct thermal field measurement is infeasible. The authors position this work as a step forward in bridging simulation-based learning with experimental analysis, enabling the inference of otherwise inaccessible thermal information from visual data.

The authors maintain a modest stance, noting that while the framework produces qualitatively accurate inferences and captures key physical trends, its quantitative performance is currently bottlenecked by the diversity of available training datasets. They assert that the approach provides a viable alternative for generating thermal data without requiring velocity maps or other difficult-to-collect experimental variables, but emphasize that acquiring diverse simulated and experimental datasets across different boiling regimes remains a fundamental challenge for future improvements.