🔬 physics

Analysis of Dominant Factors in Active Infrared Thermography Defect Detection Based on Physics-Constrained Surrogate Models

This study develops a physics-constrained framework combining a dimensionless heat transfer model and a neural network surrogate to identify dominant factors and enable rapid, efficient optimization of active infrared thermography defect detection parameters.

Original authors: Jianping Yu, Jiayi Wang, Weixiang Du, Wanhao Zheng, Xinyi Li

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jianping Yu, Jiayi Wang, Weixiang Du, Wanhao Zheng, Xinyi Li

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 are trying to find a hidden pocket of air inside a thick block of cheese without cutting it open. You can't see inside, so you decide to shine a bright heat lamp on the surface and watch how the heat travels through the cheese. If there's a pocket of air (a defect), the heat will get stuck or bounce back differently than in the solid cheese, creating a "hot spot" on the surface that a camera can see. This is the basic idea of Active Infrared Thermography, a high-tech way to check for cracks or holes in materials like airplane wings or bridge parts.

However, figuring out exactly how to set up the heat lamp is tricky. Should you blast it with high heat for a second? Or use a gentle heat for a long time? The paper by Jianping Yu and his team tackles this confusion by building a "virtual test lab" to understand the rules of the game.

Here is a simple breakdown of their findings:

1. Simplifying the Chaos: The "One-Way Street" Model

In the real world, heat moves in all directions (up, down, left, right). But the researchers realized that for thin materials, heat mostly travels straight down through the thickness, like water flowing down a narrow pipe. They simplified their math to treat the material as a one-dimensional stack of layers. This allowed them to strip away the confusing details and focus on the nine most important "knobs" they could turn, such as the material's properties, the size of the defect, and the settings of the heat lamp.

2. The Two Main Rules of the Game

After running thousands of virtual simulations, they discovered that only two things really matter when trying to spot a defect:

The "Traffic Jam" at the Interface (Thermal Impedance): Imagine heat waves as cars driving down a highway. When they hit a defect (like a bubble of air), it's like hitting a sudden traffic jam or a wall. The cars (heat) bounce back. The bigger the difference between the "road" (the material) and the "wall" (the defect), the bigger the traffic jam, and the easier it is to see the problem. The paper found that the physical properties of the material are the most critical factor here.
The "Low-Pass Filter" Effect (Time and Energy): Heat waves are like sound waves; high-pitched sounds (fast, short bursts of heat) get absorbed quickly by the material and never reach the defect. Low-pitched sounds (slower, longer bursts) travel deeper. The researchers found that the heat conduction process acts like a low-pass filter. To see deep defects, you need to send in enough total energy over a longer period, rather than just a quick, intense flash.

3. What Doesn't Matter Much

You might think that the wind blowing on the material (convection) or the exact thickness of the tiny air bubble would change the results drastically. The paper claims that under normal indoor conditions, wind and bubble thickness have very little effect. The main story is told by the material itself and how long you shine the light.

4. The "Crystal Ball" (The Surrogate Model)

Traditionally, if an engineer wanted to know the best settings for a new material, they would have to run a complex computer simulation that takes hundreds of seconds (or even minutes) for just one test. If they needed to test 1,000 different settings, it would take forever.

The team built a Neural Network Surrogate Model. Think of this as a "crystal ball" or a highly trained assistant.

How it works: They taught this AI assistant using the results from their 100 virtual simulations.
The Magic: Once trained, this AI can predict the result of a new test in milliseconds (about 100,000 times faster than the old method).
Accuracy: It is about 93% accurate compared to the slow, heavy-duty simulations.

5. Practical Advice for Engineers

Based on their "crystal ball," the paper offers simple rules for real-world inspections:

For materials that conduct heat well (like metal): You need a strong, powerful burst of heat.
For materials that are poor conductors (like fiberglass or plastic): You need a gentler, longer-lasting heat.
For big, fast inspections: Use high power for a short time to get results quickly.
For delicate materials: Use low power for a longer time to avoid damaging the part.

Summary

The paper doesn't invent a new camera or a new type of heat lamp. Instead, it provides a rulebook and a fast calculator. It tells engineers that the material's nature and the total energy of the heat pulse are the most important factors, while other details are less critical. Most importantly, it gives them a super-fast AI tool to figure out the perfect settings for any job, turning a process that used to take hours into one that takes a fraction of a second.

Technical Summary: Analysis of Dominant Factors in Active Infrared Thermography Defect Detection Based on Physics-Constrained Surrogate Models

Problem Statement
Active Infrared Thermography (IRT) is a widely used non-destructive testing (NDT) method for detecting material defects such as delaminations and voids. However, its practical application is hindered by the complexity of optimizing testing parameters. The thermal response signal is governed by a strongly coupled system involving multiple variables: thermophysical properties of the base material and defects, geometric parameters (depth and thickness), excitation waveform characteristics, and environmental heat exchange conditions. Current research lacks standardized guidelines for parameter settings, and the strong coupling between variables makes it difficult to establish clear optimization priorities, often forcing reliance on empirical operator judgment.

Methodology
To address these challenges, the authors developed a systematic framework combining dimensional analysis, numerical simulation, and machine learning:

Physics-Based Dimensionality Reduction: The study simplified the three-dimensional transient heat conduction problem into a one-dimensional model along the thickness direction, assuming thin plates where lateral dimensions exceed thickness. Using the Buckingham $\pi$ theorem, the original parameter space (involving over ten variables) was reduced to a nine-dimensional dimensionless space. Key dimensionless groups include the Fourier number ($Fo$), Biot number ($Bi$), ratios of thermal conductivity and diffusivity ( $k_1/k_0$ , $\alpha_1/\alpha_0$ ), and excitation parameters.
Numerical Simulation and Validation: A Finite Volume Method (FVM) with an explicit time-marching scheme was employed to solve the governing equations. Interface continuity was handled using the Harmonic Mean method for effective thermal conductivity. The model was validated against experimental data from GFRP specimens containing embedded defects (adhesive tape, PVC foam, paper), showing a correlation coefficient of 0.91 and relative errors below 0.5%.
Sensitivity and Parameter Analysis: Automated parametric sweeps and Latin Hypercube Sampling (LHS) were used to generate response curves. The study utilized Sobol-based sensitivity analysis to quantify the normalized sensitivity of the peak surface temperature difference ( $\Delta\theta_{max}$ ) to various dimensionless parameters.
Surrogate Model Development: To overcome the computational bottleneck of solving partial differential equations (PDEs) repeatedly, a Physics-Constrained Neural Network (PCNN) surrogate model was constructed. This feedforward neural network (10-5-5 architecture) was trained on 100 samples (80/20 split) with physical constraints embedded in the loss function to prevent overfitting and enhance generalization on small datasets.

Key Results

Dominant Mechanisms: Sensitivity analysis revealed that defect detectability is primarily driven by two factors, which together account for over 80% of the influence:
1. Interfacial Thermal Impedance Mismatch: The ratio of thermal properties between the defect and the matrix ( $k_1/k_0$ and $\alpha_1/\alpha_0$ ) is the most critical factor, determining the reflection and transmission of thermal waves.
2. Excitation Energy Input: The total energy delivered by the excitation source (primarily determined by the DC component $q_0$ and duration $t_p$ ) is the second most significant factor.
Negligible Factors: Under common operating conditions, convective heat transfer conditions (Biot number) and defect thickness ( $\delta/L$ ) were found to have limited influence on the peak thermal response.
Low-Pass Filtering Behavior: The study confirmed that transient heat conduction acts as a low-pass filter; high-frequency components of the excitation waveform attenuate rapidly, meaning the total energy integral (low-frequency content) is more effective for defect detection than high-frequency modulation.
Surrogate Model Performance: The physics-constrained neural network achieved a coefficient of determination ( $R^2$ ) of 0.929 on the test set, with a mean relative error of 6.71%.
Computational Efficiency: The surrogate model reduced the prediction time for a single evaluation from hundreds of seconds (traditional PDE simulation) to the millisecond scale, representing a computational acceleration factor of $10^5$ .

Significance and Claims
The paper claims to provide both a theoretical foundation and an efficient computational tool for the physics-driven optimization of active infrared thermography. By reducing the problem to a dimensionless framework, the study unifies different NDT scenarios (varying materials, sizes, and power levels) into a single analytical structure. The proposed surrogate model serves as a rapid forward solver, enabling efficient analysis of detection capabilities and sensitive intervals without the need for repeated, computationally expensive numerical simulations.

The authors emphasize that while the model does not directly invert defect parameters, it facilitates the optimization of excitation parameters and inspection strategies. Specifically, the study offers practical engineering guidelines: high-conductivity materials require higher power density, while low-conductivity materials (like GFRP) benefit from longer excitation durations to avoid damage. Conversely, rapid screening of large structures can utilize high-power, short-duration excitation. The work aims to support the standardization of IRT parameter settings, moving the field away from empirical judgment toward data-driven optimization.