Akshaya Srinivasan and colleagues from RPTU Kaiserslautern Landau have combined quantum and classical computing to improve quality control in industry. They demonstrated two hybrid quantum-classical machine learning approaches for identifying defects in aluminum TIG welding images and compared their effectiveness with traditional deep learning models. In this study, we use convolutional neural networks to simplify complex image data before processing it with quantum algorithms, specifically variational quantum circuits with variational quantum linear solvers and angular encoding. Although classical convolutional neural networks showed strong performance, the results show that a hybrid model can achieve competitive performance, suggesting a viable path to near-term quantum solutions for real-world industrial defect detection and quality assurance.
Quantum machine learning matches classical performance in identifying weld defects
Competitive performance is achieved between a hybrid quantum-classical model and a classical convolutional neural network, matching the accuracy of Model 1, a classical system that previously outperformed 13 other architectures. Achieving comparable results paves the way to exploiting quantum computing in practical applications, as previous quantum approaches have struggled to surpass established classical methods in complex image classification tasks. Both quantum models, employing a variational quantum linear solver and angular encoding, respectively, were benchmarked for defect classification in aluminum TIG welding images, a critical process in the industry that demands high structural integrity. The importance of this research lies in demonstrating that quantum computing can at least match the performance of state-of-the-art classical machine learning in certain industrially relevant applications, an important step in realizing the potential of quantum technology.
Model-1, a convolutional neural network specifically designed for welding image classification, validated these competitive results. The CNN acts as a feature extractor, reducing the dimensionality of the welding image before quantum processing and producing a feature vector that allows comparison with the quantum model. This dimensionality reduction is of great importance as it addresses the challenge of directly feeding quantum algorithms with high-resolution image data, which is often limited by qubit availability and coherence time. The variational quantum linear solver (VQLS) approach incorporates quantum kernel methods to map data into a high-dimensional “Hilbert space” to enhance the support vector machine (SVM) optimization process. Analysis of the quantum kernel condition number reveals its impact on classification stability. A well-tuned kernel is essential for stable and accurate classification and prevents numerical problems during the SVM training process. Variational quantum circuits (VQCs) encode classical features directly as angles within quantum gates, were trained with a classical optimizer for model improvement, and tested across both binary and multiclass defect identification scenarios. Choosing angular encoding allows for a natural mapping of classical data to quantum states, but requires careful calibration to ensure optimal performance.
Quantum kernel performance mirrors traditional networks optimized for weld defect detection
Automated visual inspection is essential to maintaining quality in manufacturing, especially for critical welds where hidden defects can have catastrophic consequences. In industries such as aerospace, automotive, and construction, weld integrity is paramount, and even minor defects can compromise the safety of a structure. Researchers are currently investigating whether quantum computing can provide an advantage in detecting these defects beyond matching the performance of existing systems. Like many studies in this area, this study shows parity, but not superiority, over carefully optimized classical convolutional neural networks. This benchmark system has already outperformed 13 other image analysis architectures. The challenge is not necessarily to immediately replace classical methods, but to identify scenarios where quantum algorithms can offer clear advantages, either in terms of accuracy, speed, or resource efficiency.
A detailed investigation of these approaches, especially how quantum kernels affect classification accuracy and improving angular encoding techniques, will be a valuable step toward using quantum power for complex image analysis tasks. By mapping data into a high-dimensional Hilbert space, quantum kernel methods have the potential to allow models to capture nonlinear relationships that may be missed by traditional SVMs. However, the effectiveness of this approach is highly dependent on the kernel selection and data characteristics. Optimization of the angular encoding scheme within VQC is also of great importance as it directly affects the expressivity and trainability of quantum circuits. The ability to effectively encode classical information into quantum states is a fundamental requirement for hybrid quantum-classical machine learning.
A hybrid quantum-classical approach is able to achieve comparable performance to traditional deep learning in classifying defects in aluminum TIG welding images, opening further developments to realize the full potential of quantum computing in industrial applications. The researchers investigated two different quantum models by first simplifying the image data using a convolutional neural network. One utilized a variational quantum linear solver and the other utilized angular encoding within a quantum circuit. Competitive results with established technology signify a move towards viable near-term quantum solutions for industrial quality control and pave the way for further research in quantum algorithms for image analysis. The use of aluminum TIG welding images provides a realistic and challenging test case because weld defects can be subtle and variable, and accurate detection requires advanced image analysis techniques. Future research may focus on exploring different quantum algorithms, optimizing hybrid architectures, and extending the approach to larger and more complex datasets. The ultimate goal is to develop a quantum-enhanced quality control system that can improve manufacturing efficiency, reduce costs, and enhance product safety.
This research highlights the potential for quantum machine learning to contribute to the continued advancement of automated industrial inspection. Although current quantum hardware limitations do not allow this task to clearly demonstrate quantum supremacy, achieving equivalence with a strong classical baseline is an important milestone. Investigating both VQLS and VQC approaches provides valuable insight into the strengths and weaknesses of various hybrid quantum-classical strategies. Continuing research into these technologies, in parallel with advances in quantum hardware, may ultimately lead to the development of robust and scalable quantum solutions for industrial quality control, offering benefits that match or exceed existing conventional performance.
This study demonstrated that a hybrid quantum-classical model can perform competitively with traditional convolutional neural networks in classifying defects in aluminum TIG welding images. This is important because it suggests that quantum computing may provide a viable means towards automated quality control in industrial settings. The researchers achieved this by combining two different quantum approaches, including one that utilizes classical image processing and a variational quantum linear solver. Future work will focus on exploring different quantum algorithms and optimizing hybrid architectures to further refine these techniques.
