⚛️ quantum physics

Hybrid Quantum-Classical AI for Industrial Defect Classification in Welding Images

This study demonstrates that hybrid quantum-classical machine learning models, utilizing quantum kernels and variational circuits for feature encoding, perform competitively against conventional deep learning approaches for classifying defects in industrial aluminium TIG welding images.

Original authors: Akshaya Srinivasan, Xiaoyin Cheng, Jianming Yi, Alexander Geng, Desislava Ivanova, Andreas Weinmann, Ali Moghiseh

Published 2026-04-01

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Akshaya Srinivasan, Xiaoyin Cheng, Jianming Yi, Alexander Geng, Desislava Ivanova, Andreas Weinmann, Ali Moghiseh

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

Imagine you are a master welder inspecting a long line of metal joints. Your job is to spot tiny cracks, dirt, or weak spots that could cause a car or airplane to fail later. Doing this by eye is slow and tiring. So, you hire a robot to do it.

This paper is about building a smarter, faster robot using a mix of old-school computing (what we have today) and futuristic quantum computing (the next big thing).

Here is the story of how they did it, explained simply:

1. The Problem: Too Much Data for a Tiny Brain

The robot needs to look at high-definition photos of welds. These photos are huge—like a massive library of details.

The Issue: Current quantum computers (the "futuristic brains") are like tiny, fragile babies. They can't handle a whole library of books at once. If you try to feed them a giant photo, they get overwhelmed and confused.
The Solution: They built a "translator." First, a standard, powerful computer (a CNN) looks at the photo and summarizes it. It says, "Okay, this photo has a lot of texture here, a dark line there, and some shiny spots." It turns the giant photo into a short, 63-word summary. Now, the tiny quantum baby can understand it.

2. The Three Contestants

The researchers tested three different "robots" to see which one could best tell the difference between a Good Weld and a Bad Weld (specifically looking for dirt or weak joints).

🏆 Contestant A: The Veteran (Classical CNN)

Who: A standard, highly trained AI that has seen thousands of photos.
How it works: It looks at the picture, remembers patterns, and shouts, "That's a bad weld!"
Result: It was perfect. It got 100% right. It's the reliable workhorse that never fails.

🥈 Contestant B: The Quantum Librarian (VQLS-Enhanced QSVM)

Who: A hybrid robot that uses quantum physics to organize data.
The Analogy: Imagine you have a pile of mixed-up cards. The classical robot sorts them one by one. The Quantum Librarian uses a special trick: it puts all the cards into a "superposition" (a magical state where they are all in many places at once) and asks, "Which pile does this card belong to?"
How it works: It takes the 63-word summary, turns it into a quantum state, and uses a complex math trick (VQLS) to solve a puzzle to find the best way to separate "Good" from "Bad."
Result: It did pretty well (about 97% on simple tasks, 92% on harder ones). It's smart, but the math is so complicated that it takes a long time to train, like trying to solve a Rubik's cube while riding a unicycle.

🥇 Contestant C: The Quantum Acrobat (VQC-Based Classifier)

Who: A flexible, trainable quantum circuit.
The Analogy: Think of this as a gymnast. You give it the 63-word summary, and it flips, twists, and turns (rotates) the data in a quantum space to find the perfect angle to spot the defect.
How it works: It encodes the data into quantum gates, spins them around, measures the result, and learns from its mistakes using a classical computer coach.
Result: It was the star of the show! It got nearly 99% right, almost matching the Veteran. It was fast, efficient, and handled the "hard" tasks (distinguishing three types of defects) better than the Quantum Librarian.

3. The Big Reveal

The researchers wanted to know: "Is quantum computing ready for the factory floor yet?"

The Good News: Yes! The "Quantum Acrobat" (Contestant C) performed almost as well as the best standard AI. This proves that even with today's noisy, imperfect quantum computers, we can build useful tools for industry.
The Catch: The "Quantum Librarian" (Contestant B) was too slow and complicated for now. It's like having a Ferrari engine in a car that's stuck in mud; it's powerful, but the setup is too heavy.
Why it worked so well: The defects they were looking for (dirt and weak joints) were very obvious visually. It was like finding a black cat in a white room. If the defects were harder to see, the quantum models might have struggled more.

The Takeaway

This paper is a proof of concept. It shows that we don't need to wait for perfect, futuristic quantum computers to start using them. By mixing them with our current computers (a Hybrid approach), we can already start building smarter quality control systems for factories.

In short: They taught a tiny, futuristic quantum brain to read a summary written by a smart classical brain, and together, they became excellent weld inspectors.

1. Problem Statement

Automated quality control in industrial manufacturing, particularly in welding processes (e.g., automotive, aerospace), is critical for structural safety. While Convolutional Neural Networks (CNNs) have achieved success in weld defect detection, they demand significant computational resources, especially as data dimensionality and model complexity increase.
The paper addresses the feasibility of Hybrid Quantum-Classical Machine Learning (HQML) for this task. The core challenge is adapting high-resolution industrial image data to the constraints of current Noisy Intermediate-Scale Quantum (NISQ) devices, which have limited qubit counts, gate fidelity, and circuit depth. The study aims to determine if quantum-enhanced models can compete with classical deep learning benchmarks in real-world industrial defect classification.

2. Methodology

The authors propose a hybrid pipeline where a classical CNN acts as a feature extractor, reducing high-dimensional image data into compact feature vectors suitable for quantum processing. Two distinct quantum classification approaches are benchmarked against a classical CNN baseline.

A. Data and Preprocessing

Dataset: A subset of an industrial TIG (Tungsten Inert Gas) welding dataset (Kaggle source) containing ~30,000 HDR grayscale images.
Task: A curated 3-class classification task:
1. Good Weld (Class 0)
2. Contamination (Class 2)
3. Lack of Fusion (Class 3)
Feature Extraction: A classical CNN (adapted from "Model-1" in prior literature) extracts feature vectors. The authors systematically tested feature vector lengths of 7, 15, 31, 63, and 127 to find the optimal balance between information retention and quantum circuit feasibility.

B. The Three Models

Classical CNN Baseline:
- Used as the performance benchmark.
- Trained end-to-end for classification using Adam optimizer and categorical cross-entropy loss.
- Serves as the feature extractor for the quantum models.
VQLS-Enhanced Quantum Support Vector Machine (QSVM):
- Kernel Construction: Classical feature vectors are normalized and encoded into quantum states using a parameterized feature map (rotation and entangling gates).
- Kernel Matrix: A quantum kernel matrix $K$ is computed via inner products of these states ( $K_{ij} = |\langle \tilde{x}_i | \tilde{x}_j \rangle|^2$ ), mapping data to a higher-dimensional Hilbert space.
- Solver: The SVM optimization problem ( $M\alpha = b$ ) is solved using the Variational Quantum Linear Solver (VQLS).
- Ansatz: A hardware-efficient ansatz using $R_y$ and controlled- $R_z$ gates.
- Optimization: A classical COBYLA optimizer minimizes a cost function based on the overlap between the trial solution and the target vector, utilizing Hadamard tests.
- Multiclass Strategy: One-vs-Rest (OvR) approach.
Variational Quantum Circuit (VQC) Classifier:
- Encoding: Features are dimensionally reduced (linear projection) to match the number of qubits (fixed at 4 for optimal performance) and angle-encoded using $R_y$ gates.
- Circuit: A single-layer hardware-efficient ansatz with trainable $R_y$ and entangling $CX$ gates.
- Measurement: Qubits are measured in the Pauli-Z basis to produce expectation values.
- Post-Processing: A classical linear layer with Softmax converts quantum outputs into class probabilities.
- Training: Optimized via Adam using cross-entropy loss and the parameter-shift rule for gradients.

C. Hyperparameter Tuning

The authors used Ray Tune with Population-Based Training (PBT) to optimize batch size, feature size, qubit count, and circuit depth. The optimal configuration identified was a feature size of 63 and 4 qubits with a batch size of 16.

3. Key Contributions

Real-World Industrial Application: Demonstrates a complete HQML pipeline on a high-resolution, real-world industrial welding dataset, moving beyond synthetic or small-scale datasets.
Comparative Benchmarking: Provides a rigorous comparison between a classical CNN, a Kernel-based QSVM (VQLS), and a Variational Quantum Classifier (VQC) on identical tasks.
Feature Size Analysis: Systematically evaluates how the dimensionality of classical feature vectors impacts quantum model performance, identifying a "sweet spot" (size 63) that balances information richness with circuit noise.
Algorithmic Adaptation: Adapts VQLS and VQC architectures specifically for NISQ constraints, utilizing classical preprocessing to mitigate qubit limitations.

4. Results

Experiments were conducted on binary (Good vs. Defective) and multiclass (3-class) tasks using a test set of 400 images.

Model	Binary Classification (Acc @ Size 63)	Multiclass Classification (Acc @ Size 63)	Key Observations
Classical CNN	100%	100%	Consistently perfect performance; serves as the robust upper bound.
VQC Classifier	99.7%	98.9%	Near-perfect performance; highly competitive with classical CNN. Efficient training.
VQLS-QSVM	96.8%	92.4%	Good performance but lower than VQC. Training is computationally expensive due to iterative kernel matrix construction and condition number sensitivity.

Feature Size Impact: Performance for quantum models peaked at a feature size of 63. Increasing to 127 led to slight performance degradation, likely due to overfitting or increased noise in the quantum circuit.
Multiclass Difficulty: All models showed a performance drop in the 3-class task compared to binary, but the VQC maintained the highest accuracy among quantum approaches.
Visual Cues: The study noted that specific visual features (e.g., dark lines for "Lack of Fusion") were critical for high accuracy, while reflections sometimes caused confusion between "Good" and "Contamination" classes.

5. Significance and Conclusion

Competitive Viability: The study proves that hybrid quantum-classical models can achieve competitive accuracy (within 1-2% of classical CNNs) on real industrial tasks, despite current hardware limitations.
VQC Superiority: The VQC-based classifier outperformed the VQLS-QSVM, likely due to the latter's sensitivity to the condition number of the kernel matrix and the high computational overhead of the linear solver.
NISQ Readiness: The results suggest that HQML is ready for near-term deployment in industrial quality assurance, particularly for tasks where classical models are already established, offering a pathway for future quantum advantage as hardware scales.
Future Outlook: As quantum hardware improves (lower noise, more qubits), these hybrid architectures are expected to become more scalable and potentially surpass classical limits in specific high-dimensional feature spaces.

In summary, the paper successfully bridges the gap between theoretical quantum machine learning and industrial application, demonstrating that hybrid VQC models are currently the most promising quantum approach for industrial defect classification.