⚡ electrical engineering

Non-invasive visualization of boiling from sound using a generative model

This paper introduces a generative model that reconstructs high-speed videos of boiling dynamics from acoustic emissions and static visual cues, effectively creating a non-invasive "virtual window" to visualize bubble behavior in optically inaccessible thermal systems where traditional imaging fails.

Original authors: Doyeong Lim, In-Cheol Bang

Published 2026-07-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Doyeong Lim, In-Cheol Bang

Original paper licensed under CC BY 4.0 (https://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

Imagine you're trying to watch a secret show happening inside a sealed, opaque box. You can't see inside, but you can hear it. The show is "boiling"—a chaotic dance of bubbles forming, growing, and popping on a hot surface. Usually, to see this dance, you need a high-speed camera. But in real-world machines, like the super-hot chips inside AI data centers or nuclear reactors, the boiling surface is often hidden behind metal, glass, or radiation. The camera is locked out.

So, here's the big question: Can we turn the sound of the boiling into a movie of the boiling?

That's exactly what Doyeong Lim and In-Cheol Bang from UNIST tried to do. And they found a way to build a "virtual window" using a special kind of AI.

The Problem: The "One-to-Many" Mystery

Think of the sound of boiling like a crowd cheering. If you hear a roar, you know people are excited, but you can't tell exactly who is jumping up and down or where they are standing just by listening.

The authors explain that this is the core problem. A single microphone outside the box hears the "roar" (the acoustic signal), but that sound is a mix of thousands of bubbles. If you tried to use a standard computer program to guess the video from the sound, it would get confused. Because one sound can come from many different bubble arrangements, a standard program would try to guess the "average" of all possibilities. The result? A blurry, fuzzy mess where the bubbles look like a soft, indistinct cloud. It's like trying to draw a specific person's face by averaging a thousand different faces together—you'd just get a blurry blob.

The Solution: The "Imagination Engine"

Instead of guessing the "average," the authors built a model that acts more like a creative storyteller. They call it a generative model based on "flow matching."

Here's how it works, using a playful analogy:
Imagine you have a blank canvas (a static image of the heater) and a script (the sound wave). You want to paint a movie of bubbles.

The Inputs: The AI is given three things it can actually get without opening the box:
- The raw sound wave (the script).
- A photo of the empty heater (the background).
- A simple mask showing where the heater is (the stage).
- Crucially, it does NOT need to have seen the boiling before or know the temperature. It works with just what's available in the real world.
The Magic: Instead of calculating one single "correct" answer, the model learns to sample from a cloud of plausible answers. It's like a jazz musician improvising. Given the same sound, the bubbles might pop in slightly different spots each time, but the rhythm and intensity will be right. The model generates a video that looks like a real, crisp, high-speed movie of bubbles, capturing the chaotic dance without blurring it out.

The Showdown: Who Painted the Best Picture?

The team didn't just build one model; they built a whole gallery of 23 different AI artists to see who could turn sound into video best. They tested:

Diffusion models (like the ones that make AI art).
Adversarial networks (where two AIs compete to make the best image).
Transformers (the big brains behind modern AI).
And their own Flow-Matching model.

They measured the results using a metric called FVD (Fréchet Video Distance), which checks how much the generated video feels like a real video, especially how the bubbles move over time.

The Winner: The authors' no-prior residual conditional flow-matching model took the top spot with a score of 109.84.

The next best was a "Diffusion Transformer" at 176.48.
The "MM-Diffusion" model came in at 224.63.

The paper explicitly rules out the idea that a simple "average" or a deterministic model (one that tries to find a single perfect answer) works here. Those models produced blurry, unconvincing videos. The paper also argues against using complex audio processing (like turning sound into a frequency image) as the input; they found that feeding the model the raw sound amplitude (the actual voltage wave) worked better because it preserved the true intensity of the boiling.

What the Model Actually Does (and Doesn't Do)

It's important to know what this "virtual window" can and cannot do.

It does: It generates a video that is temporally consistent. If you watch the video, the bubbles grow, merge, and pop in a way that matches the sound perfectly. It captures the physics of the boiling intensity.
It does NOT: It does not reconstruct the exact position of every single bubble at every millisecond. The paper is clear: because the sound is a mix, the exact location of a specific bubble is unknowable from sound alone. The model generates a plausible realization, a "what-could-be" video that is physically accurate but not a pixel-perfect replay of the exact moment.

The Results in Action

When they tested the model across different heat levels—from a gentle 40 kW m⁻² to a fierce 895 kW m⁻²—the model got better as the boiling got more intense.

At low heat, it showed isolated bubbles.
At high heat, it showed dense, coalesced vapor structures, just like the real thing.
They even tested it on "virtual" heaters (shapes and sizes the AI had never seen before). The model correctly confined the boiling to the heated area, showing it understood the rules of the game, not just memorized the pictures.

The Catch

The paper is honest about its limits. Right now, this is an offline process. It takes about 6.9 seconds to generate an 8-frame clip (which is just 80 milliseconds of video) on a single powerful graphics card. So, it's not yet a real-time window you could watch live on a screen while the machine runs. It's a tool for analysis, not a live broadcast.

The Bottom Line

The authors have shown that you can turn the "noise" of boiling into a "movie" without ever needing a camera inside the box. By using a generative model that embraces uncertainty rather than fighting it, they created a method that produces the most faithful, high-speed video of boiling from sound alone. It's not a magic trick that reveals the exact secret of every bubble, but it's the closest we've come to a "virtual high-speed window" into the hidden, churning heart of boiling systems.

Technical Summary: Non-invasive Visualization of Boiling from Sound using a Generative Model

Problem Statement
Boiling heat transfer is critical for the thermal management of high-power systems, ranging from nuclear reactors to AI data centers. The performance and safety limits of these systems are governed by rapid, near-wall bubble dynamics (nucleation, growth, coalescence, and departure). While high-speed imaging is the standard diagnostic for resolving these dynamics, most practical applications involve sealed, opaque, or optically inaccessible surfaces (e.g., immersion-cooled servers or closed cold plates), rendering direct visualization impossible.

Although acoustic emission (AE) sensing can non-invasively capture boiling intensity and regime transitions, existing data-driven methods only convert sound into scalar metrics (e.g., heat flux or regime classification). The central challenge addressed in this work is the acoustic-to-video inverse problem: converting a single external acoustic waveform into a high-speed video of the boiling process. This problem is fundamentally ill-posed and "one-to-many" because a single acoustic channel aggregates spatial information, making it impossible to uniquely determine the instantaneous pixel-level position of every bubble. Deterministic regression models trained on such data typically collapse to a blurred conditional average, failing to capture the stochastic nature of bubble dynamics.

Methodology
The authors propose a no-prior residual conditional flow-matching framework to generate stochastic boiling videos from acoustic inputs.

Data Acquisition: Experiments were conducted on a transparent pool-boiling facility with a 15 × 32 mm² Indium-Tin-Oxide (ITO) heater. Simultaneous data was collected using a high-speed camera (100 fps) and a broadband acoustic emission sensor (1 MHz sampling). The dataset spans heat fluxes from ~40 kW m⁻² to ~895 kW m⁻², covering isolated, coalesced, and vigorous boiling regimes.
Input Formulation: The model operates under a "no-prior" constraint, meaning it requires no video-derived nucleation maps, future frames, or thermal labels (heat flux/HTC) during inference. The inputs are:
- A raw-amplitude acoustic waveform (CSV).
- A static background image (pre-heating frame).
- A binary heating region-of-interest (ROI) mask.
Residual Video Generation: Instead of generating the full video frame, the model predicts the dynamic residual ( $R_{1:T} = I_{1:T} - b_0$ ), where $b_0$ is the static background. This concentrates model capacity on bubble dynamics and prevents the model from wasting capacity reconstructing static scenes.
Generative Architecture: The core is a conditional flow-matching U-Net. Unlike diffusion models that iteratively denoise, flow matching learns a continuous velocity field that transports Gaussian noise to the data distribution along a straight path.
- The acoustic signal is encoded into tokens via a 1D convolutional encoder.
- The U-Net is conditioned on these audio tokens, the background image, and the ROI mask.
- The network is trained to regress the target velocity ( $u_t = z_1 - z_0$ ), where $z_1$ is the residual video and $z_0$ is Gaussian noise.
- During inference, the model samples a single plausible realization of the residual distribution via explicit Euler integration, avoiding the "blurred average" of deterministic models.

Key Contributions

Formulation of Acoustic-to-Video Boiling: The paper establishes the first framework for generating high-speed boiling videos solely from external acoustic emissions, addressing the severe ill-posedness of the inverse problem.
Stochastic Flow Matching: The authors demonstrate that a stochastic generative approach (flow matching) is physically appropriate for this one-to-many mapping, successfully sampling plausible bubble fields rather than collapsing to a mean.
No-Prior Deployability: The model is designed for real-world deployment, requiring only inputs obtainable without optical access to the boiling surface (raw audio, static background, and heater geometry).
Comprehensive Benchmarking: The proposed model was benchmarked against 23 configurations across five generative families: Diffusion (DDPM), Adversarial (GAN/PatchGAN), Latent Diffusion, Diffusion Transformers (DiT), and Flow Matching.

Results

Performance Metrics: The proposed raw-amplitude residual flow-matching U-Net (128 base channels) achieved the lowest Fréchet Video Distance (FVD) of 109.84, significantly outperforming the next best models (MM-Diffusion at 224.63 and PatchGAN at 189.36).
Temporal vs. Spatial Fidelity: The study highlights a critical metric inversion: models with the lowest Frame-level FID (image realism) often had the worst FVD (temporal consistency). The proposed model excels at preserving temporally evolving boiling dynamics, whereas deterministic or adversarial baselines produced blurred plumes or temporally inconsistent artifacts.
Qualitative Fidelity: The model accurately reproduces the heat-flux-dependent evolution of the bubble field, from isolated bubbles at low flux to dense, coalesced vapor structures at high flux. It maintains frame-to-frame coherence and aligns bubble growth events with the acoustic input.
Ablation Studies: Removing any of the deployable conditions (background, ROI, or residual formulation) degraded video-level fidelity. Crucially, preserving the raw acoustic amplitude (rather than normalizing per chunk) was essential for conveying boiling intensity.
Generalization: The model demonstrated condition-following behavior on unseen heater geometries and camera angles, generating boiling only within the prescribed heating regions, though these were qualitative stress tests rather than quantitative validations of physical generalization.

Significance and Claims
The paper claims that non-invasive acoustic-to-video generation provides a "virtual high-speed window" into boiling surfaces that cannot be directly imaged. The primary significance lies in shifting the paradigm from scalar thermal diagnosis to stochastic visualization of physical fields.

The authors modestly define the model's capability: it generates plausible, temporally evolving boiling videos that reduce conditional-average blur, rather than reconstructing the exact instantaneous bubble field (which is physically impossible from a single acoustic channel). The work establishes that flow matching is the superior architecture for this specific physical inverse problem, offering a pathway to monitor critical heat flux (CHF) and boiling regimes in sealed, high-power systems like AI data centers and nuclear components where optical access is unavailable. The current implementation is offline (approx. 7 seconds per 8-frame chunk), and the authors note that future work is required for real-time acceleration and validation on diverse, unseen facilities.