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

arXiv:2505.00823 · cs.LG, physics.app-ph · Submitted 2025-05-01 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Bubble2Heat: Optical to Thermal Inference in Pool Boiling Using Physics-encoded Generative AI".

Jane: The paper was written by the authors from.

Tom: Stay tuned as we take you through the paper and discuss its implications.

The Core Mechanism: Tom: So, Jane, before we get into how they do it, can you explain what the "Bubble2Heat" approach is trying to solve simply?

Jane: Essentially, scientists have a huge problem measuring heat in boiling water because the flow is so chaotic and rapidly changing that traditional sensors don't give us a full picture.

Lu: They are trying to reconstruct the entire thermal field—the temperature everywhere—just by looking at what we can easily see: the bubbles themselves.

Meng: That’s where their methodology comes in, taking those high-speed images and turning them into a roadmap for inference.

Lalam: It’ moving from passively observing phenomena to actively inferring the underlying physical reality of thermal distribution.

The Core Mechanism (cont.): Tom: That's a huge shift. They use a hybrid approach, right?

Jane: Exactly; they use a physics-encoded Conditional Generative Adversarial Network, or PECGAN, to learn how the geometry relates to the temperature.

Lu: The generative model is trained on simulation data first—the LB-FD simulations—to teach it the underlying physical rules of boiling.

Meng: They use sophisticated preprocessing, like Mask R-CNN, to pull those phase contours out of experimental images so they match the structure of their simulated training data.

Lalam: It’s a powerful way to show that we don't need intrusive sensors if we can teach an AI to interpret the visual language of physics.

Improvements and Methodology: Tom: The paper highlights some clever ways they improved upon existing models, like using data augmentation.

Jane: That makes sense; since physical constraints are hard to enforce in the real world, adding those extra variations helps make the model more robust when things get weird.

Lu: I love that they're incorporating a sequence of past phase contours into the input stack, not just one image, which is a really creative way to give the AI temporal context.

Meng: The "many-to-one" inference scheme mentioned in the methods section is particularly interesting for real-world use cases; it handles ambiguity better than a simple one-to-one mapping.

Lalam: This ensures that when we move from simulated environments to experimental ones, the complexity of reality doesn's break the AI's ability to give us a coherent answer.

Results and Discussion: Tom: Now, let’s talk about the results; they achieved some impressive accuracy in their simulations.

Jane: The data shows that as they improved the model—say, increasing p, which is that input count—the error rates dropped significantly across different areas of the fluid.

Lu: But it's fascinating that they also noted a trade-off: while adding more context reduced inference error, it increased local noise in the inferred temperature gradient.

Meng: That’s an important engineering warning; we have to balance accuracy with stability, or even having more input data can make the model less predictable locally.

Lalam: This entire section shows that as a path toward universal modeling, we're starting to see how these complex physical systems behave under different conditions.

Conclusion and Wrap-up: Tom: So, summarizing all this incredible work, "Bubble2Heat: Optical to Thermal Inference in Pool Boiling Using Physics-encoded Generative AI" offers a powerful new path for analyzing heat transfer.

Jane: It’s a truly impressive way to move from visual observation to inferring the full thermal reality of boiling without relying on complex, intrusive sensors.

Lu: I think the future is massive now that they've demonstrated this; we are going to see how this scales into even more complex multiphase systems.

Meng: For me, it represents a viable engineering tool that has a clear path toward practical application in demanding thermal management designs.

Lalam: This allows us to gain deeper physical insight into these dynamic processes, leading to more predictable and efficient technologies for the entire world.

cs.LG, physics.app-ph

Submitted: 2025-05-01

Updated: 2026-02-01

Comments: 25 pages, 6 figures, supplemental information

DOI: 10.1016/j.aitf.2026.100043

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: The paper introduces a novel framework for "Optical to Thermal Inference in Pool Boiling Using Physics-encoded Generative AI," addressing the critical need to accurately estimate temperature fields

Key concepts

Bubble2Heat Approach: This methodology aims to reconstruct the entire thermal field—the temperature everywhere—in boiling water. Instead of relying on traditional sensors that fail due to chaotic flow, it uses high-speed images of bubbles themselves as input. The goal is to actively infer the underlying physical reality of thermal distribution by observing visual phenomena.
PECGAN: A Physics-encoded Conditional Generative Adversarial Network (PECGAN) is used to learn how the geometry of bubbles relates to temperature. The generative model is initially trained on simulation data (LB-FD simulations) to teach it the fundamental physical rules governing boiling, enabling interpretation of visual language.
Many-to-One Inference: This method handles ambiguity better than a simple one-to-one mapping when moving from simulated environments to real ones. It allows the AI to process multiple inputs into a single coherent output, ensuring that the complexity of reality does not break the model's ability to provide an accurate answer.

Terminology

Data Augmentation — 15× in this paper · explained in 7 episodes
Accuracy — 7× in this paper · explained in 1 episode
Generative Model — 5× in this paper · explained in 2 episodes
Physical Constraints — 5× in this paper · explained in 7 episodes
Quantization — 5× in this paper · explained in 12 episodes
Conditional Generative Adversarial Network — 4× in this paper
Consistency — 4× in this paper · explained in 3 episodes
Generative AI — 4× in this paper · explained in 2 episodes
Generative Models — 4× in this paper · explained in 2 episodes
Robustness — 4× in this paper · explained in 5 episodes
Adam — 3× in this paper · explained in 2 episodes
Core — 3× in this paper · explained in 2 episodes
Generalizability — 3× in this paper · explained in 6 episodes
Instance Segmentation — 3× in this paper · explained in 3 episodes
Fidelity — 2× in this paper · explained in 2 episodes
Generalization — 2× in this paper · explained in 13 episodes
Benchmark — 1× in this paper · explained in 2 episodes
Feature Extraction — 1× in this paper · explained in 2 episodes
Generative Adversarial Networks (GANs) — 1× in this paper · explained in 3 episodes
Human-in-the-Loop — 1× in this paper · explained in 3 episodes
Hybrid Approach — 1× in this paper · explained in 4 episodes
Inverse Problem — 1× in this paper · explained in 2 episodes
Loss Function — 1× in this paper · explained in 2 episodes
Phase Transitions — 1× in this paper · explained in 2 episodes
Physics-Informed Neural Networks (PINNs) — 1× in this paper · explained in 6 episodes
Predictive Modeling — 1× in this paper · explained in 2 episodes

Summary

The paper introduces a novel framework for Optical to Thermal Inference in Pool Boiling Using Physics-encoded Generative AI, addressing the critical need to accurately estimate temperature fields from visual bubble patterns captured during boiling. This capability is vital because direct thermal measurements are often difficult or impossible in real-world boiling systems, making the ability to infer underlying physics parameters from readily available optical data a significant advancement for heat transfer research and industrial safety.

Data Preprocessing and Input Matching

The initial steps involve rigorous preparation of experimental datasets to ensure compatibility with simulation models. Images are processed through several stages, including Phase contour extraction and accounting for Thermocouple arrangement and heating variations. A critical aspect is matching the pixel dimensions between experimental and simulation data. The required dimension X to ensure physical unit consistency after resizing is calculated by the formula:

X= l 0 over 256(x LB) 256 over(l 0, L B)

where l 0 is the characteristic length of the simulation. When matching pixel size requires a resolution greater than 1024, the authors employ padding techniques. Specifically, we instead pad the image to 2048 x 2048 with liquid cells at saturation temperature after cropping so that sufficient pixels are available for resizing. Following this, a solid layer is appended after resizing with a thickness of 5 pixels and spanning the full length of 256 pixels.

Conditional GAN Architecture and Implementation

The core inference mechanism relies on a Conditional Generative Adversarial Network (GAN). The generator and discriminator structures are detailed in Figure S3.

Generator Structure: Each 2D convolution layer utilizes a kernel size of 5 times 5, incorporates batch normalization, and employs leaky ReLU activation, with the final layer using a hyperbolic tangent activation. Paddings are used to preserve dimension and no bias term.
Discriminator Structure: The discriminator follows a similar pattern but uses varying kernel sizes: 7 times 7, 5 times 5, 3 times 3, 5 times 5 for each layer, respectively.

The model utilizes an up-sampling layer followed by a stride-1 2D convolution to mitigate artifacts, as this method is preferred to avoid checkerboard artifacts that result from 2D transposed convolution [4].

Training Protocol and Loss Functions

The training regimen is highly controlled to ensure stable learning. Each model is trained with a batch size of 1, employing the Adam optimizer with a learning rate of 0.0001. The authors note that the adversarial losses should fluctuate about 2 to ensure continuous learning. Due to an observed imbalance in training speed, the protocol dictates that we train the generator twice at each step instead of one because the discriminator usually learns faster than the generator.

Model Sensitivity and Inference Results

The model's robustness is tested by analyzing its sensitivity to variations in input conditions using dataset E2. The inference heatmaps demonstrate how changes in reference temperature (T ref) and zoom factor (z) affect the inferred thermal state.

Fig. S4 presents the Heatmap of the average inferred solid temperature.
Fig. S5 shows the Heatmap of the average inferred liquid temperature.
Fig. S6 illustrates the Heatmap of the average inferred vapor temperature.

These heatmaps confirm that variation in input parameters, particularly at large zoom factors (e.g., z=2.5 to z=5), significantly impacts the inference results across all three phases, providing a detailed mapping of model dependency on input fidelity.

Improvements for AI systems

This analysis assumes the goal is to create a more robust, generalizable, and interpretable AI system for predicting and understanding boiling heat transfer dynamics from visual/experimental data.

Here are the specific improvements that can be made to the current AI systems:

Current System: Mask R-CNN generates 8-bit instance-specific masks for bubble detection/segmentation.

Improvement: Integrate a Multi-Scale, Multi-Task Transformer Architecture (e.g., adapting methods like DETR or Swin Transformer) into the segmentation pipeline.

Fusion: Instead of relying solely on intensity values (1–256), fuse the initial mask output with supplementary physical measurements available in the dataset (e.g., localized temperature gradients from nearby thermocouples, or high-frequency acoustic emission data) as auxiliary channels.
Uncertainty Quantification: Implement Bayesian Deep Learning techniques (e.g., Monte Carlo Dropout during inference) to generate an associated predictive uncertainty map alongside the segmentation mask.

Improved AI System Capabilities:

Robust Segmentation: The system can provide highly accurate bubble masks even when contrast is low or bubbles overlap significantly, by leveraging complementary physical data streams.
Reliability Assessment: Crucially, it outputs a quantifiable measure of confidence (the uncertainty map). When the predicted uncertainty exceeds a predefined threshold (tau), the system flags the region for mandatory human review or suggests alternative physical models are required, minimizing catastrophic failure in real-world deployment.

Abstract

Phase change process plays a critical role in thermal management systems, yet quantitative characterization of multiphase heat transfer remains limited by the challenges of measuring temperature fields in chaotic, rapidly evolving flow regimes. While computational methods offer temperature data at a high spatiotemporal resolution in ideal cases, replicating complex experimental conditions remains prohibitively difficult. In this paper, we present a deep learning framework that can generate temperature field data at simulation resolution from segmented high-speed recordings and pointwise thermocouple readings which are typically available in a canonical pool boiling experimental configuration without requiring advanced techniques. This framework leverages a conditional generative adversarial network trained only on simulation data. To ensure direct applicability of the model to experimental data, our framework also introduces a preprocessing pipeline that aligns high resolution simulation data with experimental measurements through both conventional image processing and image segmentation with pretrained convolutional neural network. We further show that standard data augmentation strategies are effective in enhancing the physical plausibility of the inference when precise physical constraints are not applicable. Our results highlight the potential of deep generative models to bridge the gap between observable multiphase phenomena and underlying thermal transport, offering a powerful approach to augment and interpret experimental measurements in complex two-phase systems.